admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-06-18 16:57:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'main' into fix/fix-and-maintain-examples

Browse Source 
Signed-off-by: Nikodem Bartnik <39432165+NikodemBartnik@users.noreply.github.com>
This commit is contained in:
Nikodem Bartnik
2026-05-21 20:03:25 +02:00
committed by GitHub
parent 1451447d93 bac4f61eae
commit c2ecfc0d2b
111 changed files with 7039 additions and 1106 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/source/_toctree.yml
+10 -2
View File
@@ -3,6 +3,8 @@
title: LeRobot
- local: installation
title: Installation
- local: cheat-sheet
title: Cheat sheet
title: Get started
- sections:
- local: il_robots
@@ -37,8 +39,12 @@
title: Porting Large Datasets
- local: using_dataset_tools
title: Using the Dataset Tools
- local: dataset_subtask
title: Using Subtasks in the Dataset
- local: language_and_recipes
title: Language Columns and Recipes
- local: tools
title: Tools
- local: video_encoding_parameters
title: Video encoding parameters
- local: streaming_video_encoding
title: Streaming Video Encoding
title: "Datasets"
@@ -139,6 +145,8 @@
title: OMX
- local: openarm
title: OpenArm
- local: rebot_b601
title: reBot B601-DM
title: "Robots"
- sections:
- local: phone_teleop
docs/source/cheat-sheet.mdx
+139
View File
@@ -0,0 +1,139 @@
# Cheat sheet
All of the LeRobot commands in one place. If you forgot how to use a specific command or want to learn about a new one you can do it here.
> [!WARNING]
> For all of the commands listed below remember to change the ports/names/ids to your own values!
> [!TIP]
> Another great way to look at all the commands and get them configured for your specific setup is to use this [Jupyter Notebook](https://github.com/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb).
### Setup and installation
For installation please look at [LeRobot Installation](https://huggingface.co/docs/lerobot/main/en/installation).
### Useful tools
###### Find port
Use this to identify which serial ports your robots are connected to. Follow the instructions in your terminal: you will be asked to unplug the USB cable and press Enter. The script will then detect and print the correct serial port for that robot.
```bash
lerobot-find-port
```
###### Find cameras
Quickly find camera indices and verify their output. This command prints camera information to the terminal and saves test frames from each detected camera to `lerobot/outputs/captured_images`
```bash
lerobot-find-cameras
```
### Calibration
In most cases you will need to perform calibration just once for each robot and teleoperation device. Before performing the calibration make sure that all the joints are roughly in the middle position.
```bash
lerobot-calibrate \
--robot.type=so101_follower \
--robot.port=/dev/ttyACM0 \
--robot.id=my_follower_arm
```
Make sure that you use the same IDs used during calibration later for the other scripts. That's how LeRobot finds the calibration files.
### Teleoperation
Teleoperating with two cameras and displaying the data with Rerun.
```bash
lerobot-teleoperate \
--robot.type=so101_follower \
--robot.port=/dev/ttyACM0 \
--robot.id=my_follower_arm \
--robot.cameras="{ top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30} }" \
--teleop.type=so101_leader \
--teleop.port=/dev/ttyACM1 \
--teleop.id=my_leader_arm \
--display_data=true
```
### Recording a dataset
The dataset is automatically uploaded to the server and saved under repo_id, make sure you are logged in to your HF account with CLI:
`hf auth login`
You can get the token from: [https://huggingface.co/settings/tokens](https://huggingface.co/settings/tokens)
```bash
lerobot-record \
--robot.type=so101_follower \
--robot.port=/dev/ttyACM0 \
--robot.id=my_follower_arm \
--robot.cameras="{ top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30} }" \
--teleop.type=so101_leader \
--teleop.port=/dev/ttyACM1 \
--teleop.id=my_leader_arm \
--dataset.repo_id=${HF_USER}/so101_dataset_test \
--dataset.num_episodes=30 \
--dataset.single_task="put the red brick in a bowl" \
--dataset.streaming_encoding=true \
--display_data=true
```
While collecting the dataset you can control the process with your keyboard:
Control the data recording flow using keyboard shortcuts:
- Press **Right Arrow (`→`)**: Save episode and move to the next.
- Press **Left Arrow (`←`)**: Delete current episode and retry.
- Press **Escape (`ESC`)**: Stop, encode videos, and upload.
### Training
Depending on your hardware training the policy might take a few hours. That's how you train simple `ACT` policy:
```bash
lerobot-train \
--dataset.repo_id=${HF_USER}/so101_dataset_test \
--policy.type=act \
--output_dir=outputs/train/act_so101_test \
--job_name=act_so101_test \
--policy.device=cuda \
--wandb.enable=true \
--policy.repo_id=${HF_USER}/policy_test \
--steps=20000
```
- Policy Types: `act`, `diffusion`, `smolvla`, `pi05`
- Devices: `cuda` (NVIDIA), `mps` (Apple Silicon), `cpu`
If you want to fine-tune a specific model you can provide the path to the model. In this case path is enough and type can be skipped.
```bash
lerobot-train \
--dataset.repo_id=${HF_USER}/so101_dataset_test \
--policy.path=username/the_policy_to_finetune \
--policy.device=cuda \
--policy.repo_id=${HF_USER}/policy_test \
--output_dir=outputs/train/act_so101_test \
--steps=20000
```
### Inference
Inference means running the trained policy/model on a robot. For that we use `lerobot-rollout`. You will need to provide a path to your policy. It can be a local path or a path to Hugging Face for example "lerobot/folding_latest". Your cameras configuration needs to match what was used when collecting the dataset. Duration is in seconds if unspecified, it will run forever.
> [!TIP]
> If you are using the previous release V0.5.1 instead of `lerobot-rollout` you need to use `lerobot-record`. More information [here](https://huggingface.co/docs/lerobot/v0.5.1/en/il_robots#run-inference-and-evaluate-your-policy).
```bash
lerobot-rollout \
--strategy.type=base \
--policy.path=${HF_USER}/my_policy \
--robot.type=so101_follower \
--robot.port=/dev/ttyACM1 \
--robot.cameras="{ up: {type: opencv, index_or_path: /dev/video1, width: 640, height: 480, fps: 30}, side: {type: opencv, index_or_path: /dev/video5, width: 640, height: 480, fps: 30}}" \
--task="Put lego brick into the transparent box" \
--duration=60
```
docs/source/dataset_subtask.mdx
-277
View File
@@ -1,277 +0,0 @@
# Using Subtasks in LeRobot Datasets
Subtask support in robotics datasets has proven effective in improving robot reasoning and understanding. Subtasks are particularly useful for:
- **Hierarchical policies**: Building policies that include subtask predictions to visualize robot reasoning in real time
- **Reward modeling**: Helping reward models understand task progression (e.g., SARM-style stage-aware reward models)
- **Task decomposition**: Breaking down complex manipulation tasks into atomic, interpretable steps
LeRobotDataset now supports subtasks as part of its dataset structure, alongside tasks.
## What are Subtasks?
While a **task** describes the overall goal (e.g., "Pick up the apple and place it in the basket"), **subtasks** break down the execution into finer-grained steps:
1. "Approach the apple"
2. "Grasp the apple"
3. "Lift the apple"
4. "Move to basket"
5. "Release the apple"
Each frame in the dataset can be annotated with its corresponding subtask, enabling models to learn and predict these intermediate stages.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/subtask-asset.png"
alt="An overview of subtask annotation showing how frames are labeled with intermediate subtask stages"
width="80%"
/>
<p>
<em>Figure: Overview of subtask annotation.</em>
</p>
**Reference:** _Subtask-learning based for robot self-assembly in flexible collaborative assembly in manufacturing_, Original Article, Published: 19 April 2022.
## Dataset Structure
Subtask information is stored in the dataset metadata:
```
my-dataset/
├── data/
│ └── ...
├── meta/
│ ├── info.json
│ ├── stats.json
│ ├── tasks.parquet
│ ├── subtasks.parquet # Subtask index → subtask string mapping
│ └── episodes/
│ └── ...
└── videos/
└── ...
```
### Subtasks Parquet File
The `meta/subtasks.parquet` file maps subtask indices to their natural language descriptions:
| subtask_index | subtask (index column) |
| ------------- | ---------------------- |
| 0 | "Approach the apple" |
| 1 | "Grasp the apple" |
| 2 | "Lift the apple" |
| ... | ... |
### Frame-Level Annotations
Each frame in the dataset can include a `subtask_index` field that references the subtasks parquet file:
```python
# Example frame data in the parquet file
{
"index": 42,
"timestamp": 1.4,
"episode_index": 0,
"task_index": 0,
"subtask_index": 2, # References "Lift the apple"
"observation.state": [...],
"action": [...],
}
```
## Annotating Datasets with Subtasks
We provide a HuggingFace Space for easily annotating any LeRobotDataset with subtasks:
**[https://huggingface.co/spaces/lerobot/annotate](https://huggingface.co/spaces/lerobot/annotate)**
After completing your annotation:
1. Click "Push to Hub" to upload your annotated dataset
2. You can also run the annotation space locally by following the instructions at [github.com/huggingface/lerobot-annotate](https://github.com/huggingface/lerobot-annotate)
## Loading Datasets with Subtasks
When you load a dataset with subtask annotations, the subtask information is automatically available:
```python
from lerobot.datasets import LeRobotDataset
# Load a dataset with subtask annotations
dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
# Access a sample
sample = dataset[100]
# The sample includes both task and subtask information
print(sample["task"]) # "Collect the fruit"
print(sample["subtask"]) # "Grasp the apple"
print(sample["task_index"]) # tensor(0)
print(sample["subtask_index"]) # tensor(2)
```
### Checking for Subtask Support
You can check if a dataset has subtask annotations:
```python
# Check if subtasks are available
has_subtasks = (
"subtask_index" in dataset.features
and dataset.meta.subtasks is not None
)
if has_subtasks:
print(f"Dataset has {len(dataset.meta.subtasks)} unique subtasks")
print("Subtasks:", list(dataset.meta.subtasks.index))
```
## Using Subtasks for Training
### With the Tokenizer Processor
The `TokenizerProcessor` automatically handles subtask tokenization for Vision-Language Action (VLA) models:
```python
from lerobot.processor import TokenizerProcessorStep
# Create a tokenizer processor step
tokenizer_processor = TokenizerProcessorStep(
tokenizer_name_or_path="google/paligemma-3b-pt-224",
padding="max_length",
max_length=64,
)
# The processor will automatically tokenize subtasks if present in the batch
# and add them to the observation under:
# - "observation.subtask.tokens"
# - "observation.subtask.attention_mask"
```
When subtasks are available in the batch, the tokenizer processor adds:
- `observation.subtask.tokens`: Tokenized subtask text
- `observation.subtask.attention_mask`: Attention mask for the subtask tokens
### DataLoader with Subtasks
```python
import torch
from lerobot.datasets import LeRobotDataset
dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=16,
shuffle=True,
)
for batch in dataloader:
# Access subtask information in the batch
subtasks = batch["subtask"] # List of subtask strings
subtask_indices = batch["subtask_index"] # Tensor of subtask indices
# Use for training hierarchical policies or reward models
print(f"Batch subtasks: {set(subtasks)}")
```
## Example Datasets with Subtask Annotations
Try loading a dataset with subtask annotations:
```python
from lerobot.datasets import LeRobotDataset
# Example dataset with subtask annotations
dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
# Explore the subtasks
print("Available subtasks:")
for subtask_name in dataset.meta.subtasks.index:
print(f" - {subtask_name}")
# Get subtask distribution
subtask_counts = {}
for i in range(len(dataset)):
sample = dataset[i]
subtask = sample["subtask"]
subtask_counts[subtask] = subtask_counts.get(subtask, 0) + 1
print("\nSubtask distribution:")
for subtask, count in sorted(subtask_counts.items(), key=lambda x: -x[1]):
print(f" {subtask}: {count} frames")
```
## Use Cases
### 1. Hierarchical Policy Training
Train policies that predict both actions and current subtask:
```python
class HierarchicalPolicy(nn.Module):
def __init__(self, num_subtasks):
super().__init__()
self.action_head = nn.Linear(hidden_dim, action_dim)
self.subtask_head = nn.Linear(hidden_dim, num_subtasks)
def forward(self, observations):
features = self.encoder(observations)
actions = self.action_head(features)
subtask_logits = self.subtask_head(features)
return actions, subtask_logits
```
### 2. Stage-Aware Reward Modeling (SARM)
Build reward models that understand task progression:
```python
# SARM predicts:
# - Stage: Which subtask is being executed (discrete)
# - Progress: How far along the subtask (continuous 0-1)
class SARMRewardModel(nn.Module):
def forward(self, observations):
features = self.encoder(observations)
stage_logits = self.stage_classifier(features)
progress = self.progress_regressor(features)
return stage_logits, progress
```
### 3. Progress Visualization
Monitor robot execution by tracking subtask progression:
```python
def visualize_execution(model, observations):
for t, obs in enumerate(observations):
action, subtask_logits = model(obs)
predicted_subtask = subtask_names[subtask_logits.argmax()]
print(f"t={t}: Executing '{predicted_subtask}'")
```
## API Reference
### LeRobotDataset Properties
| Property | Type | Description |
| --------------------------- | ---------------------- | ------------------------------------------ |
| `meta.subtasks` | `pd.DataFrame \| None` | DataFrame mapping subtask names to indices |
| `features["subtask_index"]` | `dict` | Feature spec for subtask_index if present |
### Sample Keys
When subtasks are available, each sample includes:
| Key | Type | Description |
| --------------- | -------------- | ------------------------------------ |
| `subtask_index` | `torch.Tensor` | Integer index of the current subtask |
| `subtask` | `str` | Natural language subtask description |
## Related Resources
- [SARM Paper](https://arxiv.org/pdf/2509.25358) - Stage-Aware Reward Modeling for Long Horizon Robot Manipulation
- [LeRobot Annotate Space](https://huggingface.co/spaces/lerobot/annotate) - Interactive annotation tool
- [LeRobotDataset v3.0](./lerobot-dataset-v3) - Dataset format documentation
docs/source/earthrover_mini_plus.mdx
+1 -1
View File
@@ -194,7 +194,7 @@ lerobot-record \
--dataset.single_task="Navigate around obstacles" \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
# --dataset.vcodec=auto \
# --dataset.camera_encoder.vcodec=auto \
--display_data=true
```
docs/source/groot.mdx
+1 -1
View File
@@ -124,7 +124,7 @@ lerobot-rollout\
--dataset.single_task="Grab and handover the red cube to the other arm" \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
# --dataset.vcodec=auto \
# --dataset.camera_encoder.vcodec=auto \
--policy.path=<user>/groot-bimanual \ # your trained model
--duration=600
```
docs/source/hope_jr.mdx
+2 -2
View File
@@ -232,7 +232,7 @@ lerobot-record \
--dataset.private=true \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
# --dataset.vcodec=auto \
# --dataset.camera_encoder.vcodec=auto \
--display_data=true
```
@@ -278,6 +278,6 @@ lerobot-record \
--dataset.num_episodes=10 \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
# --dataset.vcodec=auto \
# --dataset.camera_encoder.vcodec=auto \
--policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
```
docs/source/il_robots.mdx
+1 -1
View File
@@ -207,7 +207,7 @@ lerobot-record \
--dataset.num_episodes=5 \
--dataset.single_task="Grab the black cube" \
--dataset.streaming_encoding=true \
# --dataset.vcodec=auto \
# --dataset.camera_encoder.vcodec=auto \
--dataset.encoder_threads=2
```
</hfoption>
docs/source/language_and_recipes.mdx
+147
View File
@@ -0,0 +1,147 @@
# Language columns and recipes
Most LeRobot datasets ship with a single `task` string per episode — fine for
short, single-instruction skills, but not enough for the longer-horizon,
multi-modal robot policies the field is moving toward (high-level planning,
memory, interjections, VQA, tool use). To support those policies without
forking the dataset format, LeRobot extends `LeRobotDataset` with two optional
language columns and a small recipe layer that turns those rows into
chat-style training samples on the fly.
The design splits cleanly into three layers:
1. **Data in the dataset** — language annotations stored next to frames in
`data/chunk-*/file-*.parquet` as two optional columns (`language_persistent`
and `language_events`). Datasets without these columns keep their existing
behavior.
2. **Recipe** — a YAML file that declares which annotation rows to bind and
how to lay them out as chat turns (`role`, `content`, optional images,
optional tool calls). Recipes are pure config; no Python required to add a
new one.
3. **Training format** — at sample time, `RenderMessagesStep` resolves the
recipe against the per-frame annotations and emits HF-style `messages` plus
LeRobot-specific sidecars (`message_streams`, `target_message_indices`)
that policy processors consume.
This page describes each layer in turn.
## Layer 1 — language columns in the dataset
The two optional columns live next to frame data in
`data/chunk-*/file-*.parquet`:
- `language_persistent`: a list of rows broadcast across every frame in an episode for state that remains active, such as `subtask`, `plan`, and `memory`.
- `language_events`: a list of rows only on the exact frame where an event was emitted, such as `interjection`, `vqa`, and speech tool calls.
Both columns share the same row shape (event rows omit `timestamp` because the
frame the row sits on already provides it):
```text
role: string
content: string | null
style: string | null
timestamp: float32 # persistent rows only
camera: string | null # observation.images.* feature key, view-dependent rows only
tool_calls: list[Json] | null
```
The `camera` field tags rows whose `content` is grounded in a specific camera
view. Rows of view-dependent styles (`vqa` and `trace`) MUST set `camera` to
the matching `observation.images.*` feature key. Rows of every other style —
including `motion`, which describes robot-frame primitives in joint / Cartesian
terms — MUST leave `camera` as `null`. Pipeline writers and the validator
enforce this via `validate_camera_field(style, camera)`.
`meta/tasks.parquet` remains the canonical source for the task. The special `${task}` recipe binding always reads that task string and does not depend on language annotations.
### Architecture
The language stack itself has three internal modules backing layer 1:
1. `lerobot.datasets.language` defines the schema, style registry, and `column_for_style`.
2. `lerobot.datasets.language_render` resolves rows and renders messages.
3. `RenderMessagesStep` turns dataset samples into `messages`, `message_streams`, and `target_message_indices`.
`LeRobotDataset` stays recipe-agnostic. It passes `language_persistent` and `language_events` through when present, and unannotated datasets keep their existing behavior.
## Layer 2 — recipe anatomy
Recipes are YAML files backed by `TrainingRecipe` and `MessageTurn`. They
declare which annotation rows to pull (via `bindings`) and how to compose them
into chat turns (`messages`).
```yaml
messages:
- { role: user, content: "${task}", stream: high_level }
- { role: assistant, content: "${subtask}", stream: low_level, target: true }
```
A recipe can also branch into a weighted **blend** of sub-recipes. At sample
time, exactly one branch is selected deterministically from the sample index,
so different frames train different objectives (e.g. memory updates vs.
low-level execution vs. VQA) without any Python wiring.
### Temporal semantics
Persistent styles are active after emission until replaced:
- `active_at(t, style=subtask)`
- `nth_prev(style=memory, offset=1)`
- `nth_next(style=subtask, offset=1)`
Event styles only exist on their exact timestamp:
- `emitted_at(t, style=interjection)`
- `emitted_at(t, style=vqa, role=user, camera=observation.images.top)`
- `emitted_at(t, role=assistant, tool_name=say)`
Exact event matching has no tolerance window, so writers must stamp event rows with frame timestamps from the parquet data.
### View-dependent resolution
For view-dependent styles (`vqa` and `trace`), the resolver gains a
`camera=` filter parallel to `role=` and `tool_name=`. Datasets with multiple
cameras typically emit one (`vqa`, `user`) + (`vqa`, `assistant`) pair per
camera at the same timestamp; without `camera=`, those resolvers see two
matches and raise an ambiguity error. Recipes consume each camera through its
own binding plus a matching image block, e.g.
```yaml
ask_vqa_top:
bindings:
vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.top)"
vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.top)"
messages:
- role: user
stream: high_level
if_present: vqa_query
content:
- { type: image, feature: observation.images.top }
- { type: text, text: "${vqa_query}" }
- {
role: assistant,
content: "${vqa}",
stream: high_level,
target: true,
if_present: vqa,
}
```
Add one such sub-recipe per camera the dataset records.
## Layer 3 — training format
Rendered samples use HF-style chat messages plus LeRobot sidecars:
```python
sample["messages"]
sample["message_streams"]
sample["target_message_indices"]
```
The renderer does not apply a tokenizer chat template. Policy processors decide how to serialize the messages for their backbone, which keeps the same dataset usable across SmolVLA, Pi0.5, and any future VLM that expects OpenAI-style chat messages.
## Graceful absence
If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.
If an event-scoped branch is selected on a frame without the required event row, rendering returns `None`, allowing a loader to retry another sample.
docs/source/lerobot-dataset-v3.mdx
+38 -1
View File
@@ -10,6 +10,7 @@ This docs will guide you to:
- Stream datasets without downloading using `StreamingLeRobotDataset`
- Apply image transforms for data augmentation during training
- Migrate existing `v2.1` datasets to `v3.0`
- Experiment with other `LeRobotDataset` formats and implementations like Lance
## What’s new in `v3`
@@ -43,7 +44,7 @@ lerobot-record \
--dataset.num_episodes=5 \
--dataset.single_task="Grab the black cube" \
--dataset.streaming_encoding=true \
# --dataset.vcodec=auto \
# --dataset.camera_encoder.vcodec=auto \
--dataset.encoder_threads=2
```
@@ -315,3 +316,39 @@ Dataset v3.0 uses incremental parquet writing with buffered metadata for efficie
- Ensures the dataset is valid for loading
Without calling `finalize()`, your parquet files will be incomplete and the dataset won't load properly.
## Other formats and implementations
### Lance
Lance is a useful format for multimodal AI datasets, especially for large-scale training requiring high performance IO and random access.
The `lerobot-lancedb` package implements `LeRobotLanceDataset` (for JPEG images) and `LeRobotLanceVideoDataset` (for mp4 videos).
Those two storage layouts both subclass LeRobotDataset and can provide data loading speed ups.
`LeRobotLanceDataset` is a drop-in replacement for `LeRobotDataset`:
```python
from lerobot.datasets import LeRobotDatasetMetadata
from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
from lerobot_lancedb import LeRobotLanceDataset, LeRobotLanceVideoDataset
cfg = DiffusionConfig(...)
meta = LeRobotDatasetMetadata(root=local_dataset_path) # or use repo_id=... to load metadata from the Hub
delta_timestamps = {...}
# Use LeRobotLanceDataset for image datasets
dataset = LeRobotLanceDataset(
root=local_dataset_path, # or use repo_id=... to stream from the Hub
delta_timestamps=delta_timestamps,
return_uint8=True,
)
# Or use LeRobotLanceVideoDataset for video datasets:
dataset = LeRobotLanceVideoDataset(
root=local_dataset_path, # or use repo_id=... to stream from the Hub
delta_timestamps=delta_timestamps,
return_uint8=True,
)
```
Join the discussion on [Github](https://github.com/huggingface/lerobot/issues/3608) and explore the `lerobot-lancedb` documentation [here](https://lancedb.github.io/lerobot-lancedb/).
docs/source/reachy2.mdx
+2 -2
View File
@@ -161,7 +161,7 @@ lerobot-record \
--dataset.private=true \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
# --dataset.vcodec=auto \
# --dataset.camera_encoder.vcodec=auto \
--display_data=true
```
@@ -203,7 +203,7 @@ lerobot-record \
--dataset.private=true \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
# --dataset.vcodec=auto \
# --dataset.camera_encoder.vcodec=auto \
--display_data=true
```
docs/source/rebot_b601.mdx
+186
View File
@@ -0,0 +1,186 @@
# reBot B601-DM
[reBot B601-DM](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/) is an open-source, low-cost robot arm from Seeed Studio for embodied-AI and imitation learning. It comes as a **follower** arm (the `B601-DM`, a 6-DOF arm plus gripper driven by Damiao CAN motors) and a **leader** arm (the `StarArm102` / `reBot Arm 102`, driven by FashionStar UART smart servos) used to teleoperate it.
This page covers **calibration** and **teleoperation** for both single-arm and bimanual (dual-arm) setups.
<div style="display: flex; align-items: center; gap: 10px;">
<img
src="https://files.seeedstudio.com/wiki/robotics/projects/lerobot/b601dm_zeroposition.jpg"
alt="reBot B601-DM follower arm at its zero position"
width="48%"
/>
<img
src="https://files.seeedstudio.com/wiki/robotics/projects/lerobot/102_zeroposition.jpg"
alt="reBot Arm 102 leader arm at its zero position"
width="48%"
/>
</div>
_Left: the B601-DM follower at its zero position. Right: the reBot Arm 102 leader at its zero position. Images courtesy of [Seeed Studio](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/)._
## Install LeRobot 🤗
Follow our [Installation Guide](./installation), then install the reBot support:
```bash
pip install -e ".[rebot]"
```
This pulls in `motorbridge` (CAN motor control for the B601-DM follower) and `motorbridge-smart-servo` (FashionStar UART servos for the reBot Arm 102 leader).
## Registered device types
| Type | Kind |
| ------------------------ | -------------------------------------------- |
| `rebot_b601_follower` | single-arm B601-DM follower robot |
| `bi_rebot_b601_follower` | bimanual (dual-arm) follower robot |
| `rebot_102_leader` | single-arm reBot Arm 102 leader teleoperator |
| `bi_rebot_102_leader` | bimanual (dual-arm) leader teleoperator |
The bimanual types compose two single-arm instances and namespace each arm's
observation/action keys with a `left_` / `right_` prefix. Per-arm settings are
passed through nested `left_arm_config.*` / `right_arm_config.*` arguments.
## Find the USB ports
For each device, find the USB port associated with its motor bus using:
```bash
lerobot-find-port
```
<Tip warning={true}>
On Linux, remove `brltty` (`sudo apt remove brltty`) so it does not hold the
leader's USB serial port. You may also need to grant access to the serial
devices: `sudo chmod 666 /dev/ttyACM* /dev/ttyUSB*`.
</Tip>
## Calibration
Neither arm stores a persistent hardware calibration: every time it connects, the motors are re-zeroed against the pose the arm is physically holding. Calibration simply records that zero pose. When prompted, **manually move the arm to its zero position** (the default sit-down pose shown above, gripper fully closed) and press <kbd>ENTER</kbd>.
### Follower (B601-DM)
<hfoptions id="calibrate-follower">
<hfoption id="Single arm">
```bash
lerobot-calibrate \
--robot.type=rebot_b601_follower \
--robot.port=/dev/ttyACM0 \
--robot.id=follower \
--robot.can_adapter=damiao
```
</hfoption>
<hfoption id="Dual arm">
Connect the bimanual follower; calibration runs for the left arm, then the right arm.
```bash
lerobot-calibrate \
--robot.type=bi_rebot_b601_follower \
--robot.id=bi_follower \
--robot.left_arm_config.port=/dev/ttyACM0 \
--robot.left_arm_config.can_adapter=damiao \
--robot.right_arm_config.port=/dev/ttyACM1 \
--robot.right_arm_config.can_adapter=damiao
```
Per-arm calibration files are saved with `_left` / `_right` suffixes on the id.
</hfoption>
</hfoptions>
### Leader (reBot Arm 102)
<hfoptions id="calibrate-leader">
<hfoption id="Single arm">
```bash
lerobot-calibrate \
--teleop.type=rebot_102_leader \
--teleop.port=/dev/ttyUSB0 \
--teleop.id=leader
```
</hfoption>
<hfoption id="Dual arm">
```bash
lerobot-calibrate \
--teleop.type=bi_rebot_102_leader \
--teleop.id=bi_leader \
--teleop.left_arm_config.port=/dev/ttyUSB0 \
--teleop.right_arm_config.port=/dev/ttyUSB1
```
</hfoption>
</hfoptions>
## Teleoperation
Once both arms are calibrated, drive the follower with the leader. The follower talks to its CAN bus through a Damiao serial bridge (`can_adapter=damiao`, the default) or a SocketCAN adapter (`can_adapter=socketcan`). See the [OpenArm page](./openarm) for more details on the SocketCAN adapter configuration.
<hfoptions id="teleoperate">
<hfoption id="Single arm">
```bash
lerobot-teleoperate \
--robot.type=rebot_b601_follower \
--robot.port=/dev/ttyACM0 \
--robot.id=follower \
--robot.can_adapter=damiao \
--teleop.type=rebot_102_leader \
--teleop.port=/dev/ttyUSB0 \
--teleop.id=leader
```
</hfoption>
<hfoption id="Dual arm">
The bimanual leader and follower reuse the single-arm classes; each arm is
configured through nested `left_arm_config.*` / `right_arm_config.*` arguments,
so a bimanual reBot Arm 102 leader drives a bimanual B601-DM follower.
```bash
lerobot-teleoperate \
--robot.type=bi_rebot_b601_follower \
--robot.id=bi_follower \
--robot.left_arm_config.port=/dev/ttyACM0 \
--robot.left_arm_config.can_adapter=damiao \
--robot.right_arm_config.port=/dev/ttyACM1 \
--robot.right_arm_config.can_adapter=damiao \
--teleop.type=bi_rebot_102_leader \
--teleop.id=bi_leader \
--teleop.left_arm_config.port=/dev/ttyUSB0 \
--teleop.right_arm_config.port=/dev/ttyUSB1
```
</hfoption>
</hfoptions>
<Tip>
The leader and follower share the same joint names (`shoulder_pan,
shoulder_lift, elbow_flex, wrist_flex, wrist_yaw, wrist_roll, gripper`), so
leader actions map directly onto the follower.
</Tip>
If the motion of a joint is reversed, flip its sign in the leader's `joint_directions` (the gripper also carries a scale to widen its range to the follower):
```bash
lerobot-teleoperate \
--robot.type=rebot_b601_follower \
--robot.port=/dev/ttyACM0 \
--robot.can_adapter=damiao \
--teleop.type=rebot_102_leader \
--teleop.port=/dev/ttyUSB0 \
--teleop.joint_directions='{"shoulder_pan":-1,"shoulder_lift":-1,"elbow_flex":1,"wrist_flex":1,"wrist_yaw":1,"wrist_roll":-1,"gripper":-6}'
```
## Recording datasets
Swap `lerobot-teleoperate` for `lerobot-record` (with the same `--robot.*` / `--teleop.*` arguments, plus `--dataset.*`) to record demonstrations for training. See [Imitation Learning for Robots](./il_robots) for the full workflow.
For hardware assembly and wiring, see the [Seeed Studio reBot wiki](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/).
docs/source/streaming_video_encoding.mdx
+22 -22
View File
@@ -17,9 +17,9 @@ This makes `save_episode()` near-instant (the video is already encoded by the ti
| Parameter | CLI Flag | Type | Default | Description |
| ----------------------- | --------------------------------- | ------------- | ------------- | ----------------------------------------------------------------- |
| `streaming_encoding` | `--dataset.streaming_encoding` | `bool` | `True` | Enable real-time encoding during capture |
| `vcodec` | `--dataset.vcodec` | `str` | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder |
| `vcodec` | `--dataset.camera_encoder.vcodec` | `str` | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder |
| `encoder_threads` | `--dataset.encoder_threads` | `int \| None` | `None` (auto) | Threads per encoder instance. `None` will leave the vcoded decide |
| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int` | `60` | Max buffered frames per camera (~2s at 30fps). Consumes RAM |
| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int` | `30` | Max buffered frames per camera (~1s at 30fps). Consumes RAM |
## 3. Performance Considerations
@@ -48,7 +48,7 @@ This parameter controls how many threads each encoder instance uses internally:
### Backpressure and Frame Dropping
Each camera has a bounded queue (`encoder_queue_maxsize`, default 60 frames). When the encoder can't keep up:
Each camera has a bounded queue (`encoder_queue_maxsize`, default 30 frames). When the encoder can't keep up:
1. The queue fills up (consuming RAM)
2. New frames are **dropped** (not blocked) — the capture loop continues uninterrupted
@@ -82,15 +82,15 @@ Use HW encoding when:
### Available HW Encoders
| Encoder | Platform | Hardware | CLI Value |
| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------ |
| `h264_videotoolbox` | macOS | Apple Silicon / Intel | `--dataset.vcodec=h264_videotoolbox` |
| `hevc_videotoolbox` | macOS | Apple Silicon / Intel | `--dataset.vcodec=hevc_videotoolbox` |
| `h264_nvenc` | Linux/Windows | NVIDIA GPU | `--dataset.vcodec=h264_nvenc` |
| `hevc_nvenc` | Linux/Windows | NVIDIA GPU | `--dataset.vcodec=hevc_nvenc` |
| `h264_vaapi` | Linux | Intel/AMD GPU | `--dataset.vcodec=h264_vaapi` |
| `h264_qsv` | Linux/Windows | Intel Quick Sync | `--dataset.vcodec=h264_qsv` |
| `auto` | Any | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.vcodec=auto` |
| Encoder | Platform | Hardware | CLI Value |
| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | --------------------------------------------------- |
| `h264_videotoolbox` | macOS | Apple Silicon / Intel | `--dataset.camera_encoder.vcodec=h264_videotoolbox` |
| `hevc_videotoolbox` | macOS | Apple Silicon / Intel | `--dataset.camera_encoder.vcodec=hevc_videotoolbox` |
| `h264_nvenc` | Linux/Windows | NVIDIA GPU | `--dataset.camera_encoder.vcodec=h264_nvenc` |
| `hevc_nvenc` | Linux/Windows | NVIDIA GPU | `--dataset.camera_encoder.vcodec=hevc_nvenc` |
| `h264_vaapi` | Linux | Intel/AMD GPU | `--dataset.camera_encoder.vcodec=h264_vaapi` |
| `h264_qsv` | Linux/Windows | Intel Quick Sync | `--dataset.camera_encoder.vcodec=h264_qsv` |
| `auto` | Any | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.camera_encoder.vcodec=auto` |
> [!NOTE]
> In order to use the HW accelerated encoders you might need to upgrade your GPU drivers.
@@ -100,15 +100,15 @@ Use HW encoding when:
## 5. Troubleshooting
| Symptom | Likely Cause | Fix |
| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage) | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.vcodec=auto`) |
| "Encoder queue full" warnings or dropped frames in dataset | Encoder can't keep up (Queue overflow) | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.vcodec=auto`). |
| High RAM usage | Queue filling faster than encoding | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding |
| Large video files | Using HW encoder or H.264 | Expected trade-off. Switch to `libsvtav1` if CPU allows |
| `save_episode()` still slow | `streaming_encoding` is `False` | Set `--dataset.streaming_encoding=true` |
| Encoder thread crash | Codec not available or invalid settings | Check `vcodec` is installed, try `--dataset.vcodec=auto` |
| Recorded dataset is missing frames | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected. |
| Symptom | Likely Cause | Fix |
| ------------------------------------------------------------------ | -------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage) | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.camera_encoder.vcodec=auto`) |
| "Encoder queue full" warnings or dropped frames in dataset | Encoder can't keep up (Queue overflow) | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.camera_encoder.vcodec=auto`). |
| High RAM usage | Queue filling faster than encoding | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding |
| Large video files | Using HW encoder or H.264 | Expected trade-off. Switch to `libsvtav1` if CPU allows |
| `save_episode()` still slow | `streaming_encoding` is `False` | Set `--dataset.streaming_encoding=true` |
| Encoder thread crash | Codec not available or invalid settings | Check `vcodec` is installed, try `--dataset.camera_encoder.vcodec=auto` |
| Recorded dataset is missing frames | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected. |
## 6. Recommended Configurations
@@ -146,7 +146,7 @@ On very constrained systems, streaming encoding may compete too heavily with the
# 2camsx 640x480x3 @30fps: Requires some tuning.
# Use H.264, disable streaming, consider batching encoding
lerobot-record --dataset.vcodec=h264 --dataset.streaming_encoding=false ...
lerobot-record --dataset.camera_encoder.vcodec=h264 --dataset.streaming_encoding=false ...
```
## 7. Closing note
docs/source/tools.mdx
+210
View File
@@ -0,0 +1,210 @@
# Tools
LeRobot v3.1 supports **tool calls** in policies — assistant messages can
emit structured invocations like `say(text="OK, starting now")` that the
runtime dispatches to a real implementation (TTS, controller, logger, …).
This page covers:
1. Where the tool catalog lives.
2. How the annotation pipeline produces tool-call atoms.
3. How to add your own tool.
## Where tools are declared
Two layers.
**The catalog** — a list of OpenAI-style function schemas — lives at
`meta/info.json["tools"]` on each dataset. Example:
```json
{
"features": { "...": "..." },
"tools": [
{
"type": "function",
"function": {
"name": "say",
"description": "Speak a short utterance to the user via the TTS executor.",
"parameters": {
"type": "object",
"properties": {
"text": {
"type": "string",
"description": "The verbatim text to speak."
}
},
"required": ["text"]
}
}
}
]
}
```
Read it via the dataset metadata accessor:
```python
from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
meta = LeRobotDatasetMetadata(repo_id="pepijn/super_poulain_final_annotations")
tools = meta.tools # list[dict] — OpenAI tool schemas
```
If the dataset's `info.json` doesn't declare any tools, `meta.tools`
returns `DEFAULT_TOOLS` from `lerobot.datasets.language` — currently a
single-entry list with the canonical `say` schema. So unannotated
datasets and chat-template consumers keep working without any
configuration:
```python
prompt_str = tokenizer.apply_chat_template(
sample["messages"],
tools=meta.tools, # works either way
add_generation_prompt=False,
tokenize=False,
)
```
**The implementations** — runnable Python — will live under
`src/lerobot/tools/`, one file per tool. The runtime dispatcher and
the canonical `say` implementation (wrapping Kyutai's pocket-tts) are
not part of the catalog layer described here; today this layer ships
only the schema storage and the `DEFAULT_TOOLS` fallback constant.
## Per-row tool _invocations_
The catalog above describes _what can be called_. The actual _call_ — the
function name plus the argument values — is stored per-row, on the
assistant atoms in `language_events`:
```python
{
"role": "assistant",
"content": null,
"style": null,
"timestamp": 12.4,
"camera": null,
"tool_calls": [
{ "type": "function",
"function": { "name": "say", "arguments": { "text": "On it." } } }
]
}
```
Recipes splice these into rendered messages via `tool_calls_from`:
```yaml
user_interjection_response:
bindings:
speech: "emitted_at(t, role=assistant, tool_name=say)"
messages:
- { role: user, content: "${task}", stream: high_level }
- {
role: assistant,
content: "${current_plan}",
stream: high_level,
target: true,
tool_calls_from: speech,
}
```
The model's training target is one assistant turn that carries both the
plan text _and_ the `say` tool call. At inference, the runtime parses
the generated text back into structured `tool_calls` and dispatches to
the matching implementation.
## How to add your own tool
> **Note:** Steps 2 and 3 below describe the runtime layer
> (`src/lerobot/tools/`, the `Tool` protocol, `TOOL_REGISTRY`,
> `get_tools(meta)`) which is not part of the catalog layer shipped
> today — those modules don't yet exist in the tree. Step 1 alone is
> enough to make the tool visible to the chat template via
> `meta.tools` so the model can learn to _generate_ the call;
> executing the call at inference requires the runtime layer.
Three steps. Concrete example: a `record_observation` tool the policy
can call to capture an extra observation outside the regular control
loop.
### Step 1 — declare the schema
Add an entry under `meta/info.json["tools"]`. Either edit the file
directly on disk _before_ running the annotation pipeline (it'll be
preserved) or hand it to `lerobot-annotate` via a config flag.
```json
{
"tools": [
{ "type": "function", "function": { "name": "say", "...": "..." } },
{
"type": "function",
"function": {
"name": "record_observation",
"description": "Capture a high-resolution still image for the user.",
"parameters": {
"type": "object",
"properties": {
"label": {
"type": "string",
"description": "Short label for the saved image."
}
},
"required": ["label"]
}
}
}
]
}
```
The schema follows OpenAI's function-calling convention exactly, so the
chat template can render it natively.
### Step 2 — implement the call
Create `src/lerobot/tools/record_observation.py`:
```python
from .base import Tool
from typing import Any
RECORD_OBSERVATION_SCHEMA: dict[str, Any] = { "...": "..." } # mirrors the JSON above
class RecordObservationTool:
name = "record_observation"
schema = RECORD_OBSERVATION_SCHEMA
def __init__(self, schema: dict | None = None, output_dir: str = "."):
self.output_dir = output_dir
def call(self, arguments: dict) -> str:
label = arguments["label"]
# ... save the latest camera frame to <output_dir>/<label>.png ...
return f"saved {label}.png"
```
One file per tool keeps dependencies isolated — `record_observation`
might pull `pillow`, while `say` pulls `pocket-tts`. Users installing
only the tools they need avoid heavy transitive deps.
### Step 3 — register it
Add to `src/lerobot/tools/registry.py`:
```python
from .record_observation import RecordObservationTool
TOOL_REGISTRY["record_observation"] = RecordObservationTool
```
That's it. At runtime `get_tools(meta)` looks up each schema in
`meta.tools`, instantiates the matching registered class, and returns
a name → instance dict the dispatcher can route into.
If you want to use a tool _without_ writing an implementation (e.g. for
training-time chat-template formatting only), step 1 alone is enough —
the model still learns to _generate_ the call. Steps 2 and 3 are only
needed to actually _execute_ it at inference.
docs/source/using_dataset_tools.mdx
+5 -9
View File
@@ -117,10 +117,10 @@ lerobot-edit-dataset \
--repo_id lerobot/pusht_image \
--operation.type convert_image_to_video \
--operation.output_dir outputs/pusht_video \
--operation.vcodec libsvtav1 \
--operation.pix_fmt yuv420p \
--operation.g 2 \
--operation.crf 30
--operation.camera_encoder.vcodec libsvtav1 \
--operation.camera_encoder.pix_fmt yuv420p \
--operation.camera_encoder.g 2 \
--operation.camera_encoder.crf 30
# Convert only specific episodes
lerobot-edit-dataset \
@@ -147,11 +147,7 @@ lerobot-edit-dataset \
**Parameters:**
- `output_dir`: Custom output directory (optional - by default uses `new_repo_id` or `{repo_id}_video`)
- `vcodec`: Video codec to use - options: `h264`, `hevc`, `libsvtav1` (default: `libsvtav1`)
- `pix_fmt`: Pixel format - options: `yuv420p`, `yuv444p` (default: `yuv420p`)
- `g`: Group of pictures (GOP) size - lower values give better quality but larger files (default: 2)
- `crf`: Constant rate factor - lower values give better quality but larger files, 0 is lossless (default: 30)
- `fast_decode`: Fast decode tuning option (default: 0)
- `camera_encoder`: Video encoder settings — all sub-fields accessible via `--operation.camera_encoder.<field>. See [Video Encoding Parameters](./video_encoding_parameters) for more details.
- `episode_indices`: List of specific episodes to convert (default: all episodes)
- `num_workers`: Number of parallel workers for processing (default: 4)
docs/source/video_encoding_parameters.mdx
+117
View File
@@ -0,0 +1,117 @@
# Video encoding parameters
When video storage is enabled, LeRobot stores each camera stream as an **MP4** file instead of saving one image file per timestep. Video encoding compresses across time, which usually cuts dataset size and I/O compared to a pile of PNG, while keeping MP4 — a format every player and loader understands.
Encoding frames into an MP4 is a full FFmpeg pipeline: choice of encoder, pixel format, GOP/keyframes, quality vs. speed, and optional extra encoder flags. Most of these knobs are user-tunable through `camera_encoder`, a nested `VideoEncoderConfig` (`lerobot.configs.video.VideoEncoderConfig`) passed through PyAV.
You can set these parameters from the CLI with `--dataset.camera_encoder.<field>` (e.g. with `lerobot-record` or `lerobot-rollout`). The same block applies to every camera video stream in that run.
<Tip>
Video storage must be on for `camera_encoder` to have any effect —
`use_videos=True` in Python APIs, or `--dataset.video=true` on the CLI (the
recording default). With video off, inputs stay as images and `camera_encoder`
is ignored.
</Tip>
For details on **when** frames are written vs. encoded (streaming vs. post-episode), queues, and other top-level `--dataset.*` switches, see [Streaming Video Encoding](./streaming_video_encoding). For an encoding-parameter comparison and experiments, see the [video-benchmark Space](https://huggingface.co/spaces/lerobot/video-benchmark).
---
## Example
```bash
lerobot-record \
--robot.type=so100_follower \
--robot.port=/dev/tty.usbmodem58760431541 \
--robot.cameras="{laptop: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
--robot.id=black \
--teleop.type=so100_leader \
--teleop.port=/dev/tty.usbmodem58760431551 \
--teleop.id=blue \
--dataset.repo_id=<my_username>/<my_dataset_name> \
--dataset.num_episodes=2 \
--dataset.single_task="Grab the cube" \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
--dataset.camera_encoder.vcodec=h264 \
--dataset.camera_encoder.preset=fast \
--dataset.camera_encoder.extra_options={"tune": "film", "profile:v": "high", "bf": 2} \
--display_data=true
```
---
## Tuning parameters
<Tip warning={true}>
The defaults are tuned to balance **compression ratio**, **visual quality**, and **decoding/seek speed** for typical robotics datasets. Changing them can affect both recording (CPU load, frame drops) and training (decoding throughput, image quality).
Only override these parameters if you have a specific reason to, and measure the impact on your pipeline before relying on the new settings.
</Tip>
All flags below are prefixed with `--dataset.camera_encoder.` on the CLI.
| Parameter | Type | Default | Description |
| --------------- | ---------------- | ------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `vcodec` | `str` | `"libsvtav1"` | Video codec name. `"auto"` picks the first available hardware encoder from a fixed preference list, falling back to `libsvtav1`. |
| `pix_fmt` | `str` | `"yuv420p"` | Output pixel format. Must be supported by the chosen codec in your FFmpeg build. |
| `g` | `int` | `2` | GOP size — a keyframe every `g` frames. Emitted as FFmpeg option `g`. |
| `crf` | `int` or `float` | `30` | Abstract quality value, mapped per codec (see the [mapping](#mapping-videoencoderconfig--ffmpeg-options) below). Lower → higher quality / larger output where the mapping is monotone. |
| `preset` | `int` or `str` | `12` \* | Encoder speed preset; meaning depends on the codec. <br/>\* When unset and `vcodec=libsvtav1`, LeRobot defaults to `12`. |
| `fast_decode` | `int` | `0` | `libsvtav1`: `0–2`, passed via `svtav1-params`. <br/>`h264` / `hevc` (software): if `>0`, sets `tune=fastdecode`. <br/>Other codecs: usually unused. |
| `video_backend` | `str` | `"pyav"` | Only `"pyav"` is currently implemented for video encoding. |
| `extra_options` | `dict` | `{}` | Extra FFmpeg or codec specific options merged after the structured fields above. Cannot override keys already set by those fields. |
---
## Persistence in dataset metadata
After the first episode of a video stream is encoded, the encoder configuration is **persisted into the dataset metadata** (`meta/info.json`) under each video feature, alongside the values probed from the file itself. For a video feature `observation.images.<camera>`, the layout in `info.json` is:
```json
{
"features": {
"observation.images.laptop": {
"dtype": "video",
"shape": [480, 640, 3],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"video.is_depth_map": false,
"video.g": 2,
"video.crf": 30,
"video.preset": "fast",
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": { "tune": "film", "profile:v": "high", "bf": 2 }
}
}
}
}
```
Two sources contribute to the `info` block:
- **Stream-derived** (read back from the encoded MP4 with PyAV): `video.height`, `video.width`, `video.codec`, `video.pix_fmt`, `video.fps`, `video.channels`, `video.is_depth_map`, plus `audio.*` if an audio stream is present.
- **Encoder-derived** (taken from `VideoEncoderConfig`): `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.video_backend`, `video.extra_options`.
<Tip>
This block is populated **once**, from the **first** episode. It assumes every
episode in the dataset was encoded with the same `camera_encoder`. Changing
encoder settings partway through a recording is not supported — the
`info.json` will only reflect the parameters used for the first episode.
</Tip>
---
## Merging datasets
When aggregating datasets with `merge_datasets`, video files are concatenated as-is (no re-encoding), and encoder fields in `info.json` are merged per-key:
- **Stream-derived fields must match** across sources: `video.codec`, `video.pix_fmt`, `video.height`, `video.width`, `video.fps`. Otherwise FFmpeg's concat demuxer fails.
- **Encoder-tuning fields are merged loosely**: `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.extra_options`. If every source agrees, the value is kept; if not, it's set to `null` (or `{}` for `video.extra_options`) and a warning is logged.
examples/dataset/create_progress_videos.py
+37 -15
View File
@@ -15,10 +15,12 @@
# limitations under the License.
"""
Create MP4 (or GIF) videos with sarm_progress overlay for specified episodes.
Create MP4 (or GIF) videos with per-frame progress overlay for specified episodes.
Downloads datasets from HuggingFace, seeks directly into the episode segment
of the source video, draws a progress line on each frame, and writes the result.
The progress data is read from a parquet file that lives alongside the dataset
(configurable via ``--progress-file``).
Usage:
python examples/dataset/create_progress_videos.py \
@@ -56,22 +58,26 @@ SCORE_FONT_SCALE = 0.8
TASK_FONT_SCALE = 0.55
def download_episode_metadata(repo_id: str, episode: int) -> Path:
"""Download only the metadata and sarm_progress files for a dataset.
def download_episode_metadata(
repo_id: str, episode: int, progress_file: str = "sarm_progress.parquet"
) -> Path:
"""Download only the metadata and per-frame progress file for a dataset.
Args:
repo_id: HuggingFace dataset repository ID.
episode: Episode index (used for logging only; all meta is fetched).
progress_file: Filename of the per-frame progress parquet inside the
dataset repo.
Returns:
Local cache path for the downloaded snapshot.
"""
logging.info("[1/4] Downloading metadata for %s (episode %d) ...", repo_id, episode)
logging.info("[1/4] Downloading metadata + %s for %s (episode %d) ...", progress_file, repo_id, episode)
local_path = Path(
snapshot_download(
repo_id=repo_id,
repo_type="dataset",
allow_patterns=["meta/**", "sarm_progress.parquet"],
allow_patterns=["meta/**", progress_file],
ignore_patterns=["*.mp4"],
)
)
@@ -215,25 +221,28 @@ def download_video_file(repo_id: str, local_path: Path, video_rel: str) -> Path:
return video_path
def load_progress_data(local_path: Path, episode: int) -> np.ndarray | None:
"""Load sarm_progress values for an episode.
def load_progress_data(
local_path: Path, episode: int, progress_file: str = "sarm_progress.parquet"
) -> np.ndarray | None:
"""Load per-frame progress values for an episode.
Args:
local_path: Dataset cache root.
episode: Episode index.
progress_file: Filename of the per-frame progress parquet.
Returns:
Sorted (N, 2) array of (frame_index, progress), or None if unavailable.
"""
parquet_path = local_path / "sarm_progress.parquet"
parquet_path = local_path / progress_file
if not parquet_path.exists():
logging.warning("sarm_progress.parquet not found")
logging.warning("%s not found", progress_file)
return None
df = pd.read_parquet(parquet_path)
logging.info(" sarm_progress.parquet columns: %s", list(df.columns))
logging.info(" %s columns: %s", progress_file, list(df.columns))
episode_df = df[df["episode_index"] == episode].copy()
if episode_df.empty:
logging.warning("No sarm_progress rows for episode %d", episode)
logging.warning("No progress rows for episode %d in %s", episode, progress_file)
return None
episode_df = episode_df.sort_values("frame_index")
@@ -576,6 +585,7 @@ def process_dataset(
camera_key: str | None,
output_dir: Path,
create_gif: bool = False,
progress_file: str = "sarm_progress.parquet",
) -> Path | None:
"""Full pipeline: download, extract metadata, composite progress, write output.
@@ -585,6 +595,8 @@ def process_dataset(
camera_key: Camera key to use, or None for auto-selection.
output_dir: Directory to write output files.
create_gif: If True, also generate a GIF from the MP4.
progress_file: Filename of the per-frame progress parquet inside the
dataset repo.
Returns:
Path to the final output file, or None on failure.
@@ -592,7 +604,7 @@ def process_dataset(
safe_name = repo_id.replace("/", "_")
logging.info("Processing: %s | episode %d", repo_id, episode)
local_path = download_episode_metadata(repo_id, episode)
local_path = download_episode_metadata(repo_id, episode, progress_file)
logging.info(" Local cache: %s", local_path)
episode_meta = load_episode_meta(local_path, episode, camera_key)
@@ -600,9 +612,9 @@ def process_dataset(
video_path = download_video_file(repo_id, local_path, episode_meta["video_rel"])
progress_data = load_progress_data(local_path, episode)
progress_data = load_progress_data(local_path, episode, progress_file)
if progress_data is None:
logging.error("Could not load sarm_progress data. Skipping overlay.")
logging.error("Could not load progress data from %s. Skipping overlay.", progress_file)
return None
logging.info(" Progress frames: %d", len(progress_data))
@@ -627,7 +639,7 @@ def process_dataset(
def main() -> None:
parser = argparse.ArgumentParser(
description="Create MP4/GIF videos with sarm_progress overlay for dataset episodes."
description="Create MP4/GIF videos with per-frame progress overlay for dataset episodes."
)
parser.add_argument(
"--repo-id",
@@ -658,6 +670,15 @@ def main() -> None:
action="store_true",
help="Also generate a GIF from the MP4 output.",
)
parser.add_argument(
"--progress-file",
type=str,
default="sarm_progress.parquet",
help=(
"Filename of the per-frame progress parquet inside the dataset repo "
"(default: 'sarm_progress.parquet')."
),
)
args = parser.parse_args()
logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
@@ -670,6 +691,7 @@ def main() -> None:
camera_key=args.camera_key,
output_dir=args.output_dir,
create_gif=args.gif,
progress_file=args.progress_file,
)
if result:
examples/notebooks/quickstart.ipynb
+29 -2
View File
@@ -80,7 +80,7 @@
"}\n",
"\n",
"# Dataset\n",
"HF_USER = \"your_hf_username\" # `huggingface-cli whoami` to find your username\n",
"HF_USER = \"your_hf_username\" # `hf auth whoami` to find your username\n",
"DATASET_NAME = \"my_so101_dataset\"\n",
"TASK_DESCRIPTION = \"pick and place the block\"\n",
"NUM_EPISODES = 10\n",
@@ -291,7 +291,34 @@
"\n",
"Uses `POLICY_PATH` from the Configuration cell (defaults to the Hub repo ID). You can also put there the `LAST_CHECKPOINT_PATH`.\n",
"\n",
"See the [inference docs](https://huggingface.co/docs/lerobot/il_robots#run-inference-and-evaluate-your-policy) for details."
"See the [inference docs](https://huggingface.co/docs/lerobot/il_robots#run-inference-and-evaluate-your-policy) for details.\n",
"\n",
"Recently ```lerobot-rollout``` was introduced, you can [read more about it here](https://huggingface.co/docs/lerobot/main/en/il_robots?eval=Base+mode+%28no+recording%29#run-inference-and-evaluate-your-policy)."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"print_cmd(\n",
" \"lerobot-rollout\",\n",
" \"--strategy.type=base\",\n",
" f\"--policy.path={POLICY_PATH}\",\n",
" f\"--robot.type={ROBOT_TYPE}\",\n",
" f\"--robot.port={ROBOT_PORT}\",\n",
" CAMERAS_FLAG,\n",
" f'--task=\"{TASK_DESCRIPTION}\"',\n",
" \"--duration=60\",\n",
")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"if you are using the V0.5.1 release you should use ```lerobot-record``` instead of rollout"
]
},
{
pyproject.toml
+7 -1
View File
@@ -95,7 +95,7 @@ dependencies = [
# ── Feature-scoped extras ──────────────────────────────────
dataset = [
"datasets>=4.0.0,<5.0.0",
"datasets>=4.7.0,<5.0.0",
"pandas>=2.0.0,<3.0.0", # NOTE: Transitive dependency of datasets
"pyarrow>=21.0.0,<30.0.0", # NOTE: Transitive dependency of datasets
"lerobot[av-dep]",
@@ -151,6 +151,8 @@ pyserial-dep = ["pyserial>=3.5,<4.0"]
deepdiff-dep = ["deepdiff>=7.0.1,<9.0.0"]
pynput-dep = ["pynput>=1.7.8,<1.9.0"]
pyzmq-dep = ["pyzmq>=26.2.1,<28.0.0"]
motorbridge-dep = ["motorbridge>=0.3.2,<0.4.0"]
motorbridge-smart-servo-dep = ["motorbridge-smart-servo>=0.0.4,<0.1.0"]
# Motors
feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0", "lerobot[pyserial-dep]", "lerobot[deepdiff-dep]"]
@@ -174,6 +176,9 @@ unitree_g1 = [
"lerobot[pygame-dep]",
]
reachy2 = ["reachy2_sdk>=1.0.15,<1.1.0"]
# Seeed Studio reBot B601-DM follower (motorbridge / CAN) + StarArm102 / reBot Arm 102
# leader (motorbridge-smart-servo / FashionStar UART servos).
rebot = ["lerobot[motorbridge-dep]", "lerobot[motorbridge-smart-servo-dep]"]
kinematics = ["lerobot[placo-dep]"]
intelrealsense = [
"pyrealsense2>=2.55.1.6486,<2.57.0 ; sys_platform != 'darwin'",
@@ -260,6 +265,7 @@ all = [
"lerobot[lekiwi]",
"lerobot[openarms]",
"lerobot[reachy2]",
"lerobot[rebot]",
"lerobot[kinematics]",
"lerobot[intelrealsense]",
"lerobot[diffusion]",
src/lerobot/cameras/opencv/camera_opencv.py
+9 -3
View File
@@ -199,12 +199,13 @@ class OpenCVCamera(Camera):
DeviceNotConnectedError: If the camera is not connected.
"""
# Set FOURCC first (if specified) as it can affect available FPS/resolution options
if self.config.fourcc is not None:
self._validate_fourcc()
if self.videocapture is None:
raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
set_fourcc_after_size_and_fps = platform.system() == "Windows"
if self.config.fourcc is not None and not set_fourcc_after_size_and_fps:
self._validate_fourcc()
default_width = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_WIDTH)))
default_height = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
@@ -222,6 +223,11 @@ class OpenCVCamera(Camera):
else:
self._validate_fps()
if self.config.fourcc is not None and set_fourcc_after_size_and_fps:
# On Windows with DSHOW, changing the resolution can silently override the FOURCC setting.
# Set FOURCC last to make sure the requested pixel format is actually enforced.
self._validate_fourcc()
def _validate_fps(self) -> None:
"""Validates and sets the camera's frames per second (FPS)."""
src/lerobot/configs/__init__.py
+16
View File
@@ -24,6 +24,7 @@ Import them directly: ``from lerobot.configs.train import TrainPipelineConfig``
from .dataset import DatasetRecordConfig
from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
from .policies import PreTrainedConfig
from .recipe import MessageTurn, TrainingRecipe, load_recipe
from .types import (
FeatureType,
NormalizationMode,
@@ -31,6 +32,12 @@ from .types import (
PolicyFeature,
RTCAttentionSchedule,
)
from .video import (
VALID_VIDEO_CODECS,
VIDEO_ENCODER_INFO_KEYS,
VideoEncoderConfig,
camera_encoder_defaults,
)
__all__ = [
# Types
@@ -43,7 +50,16 @@ __all__ = [
"DatasetRecordConfig",
"DatasetConfig",
"EvalConfig",
"MessageTurn",
"PeftConfig",
"PreTrainedConfig",
"TrainingRecipe",
"WandBConfig",
"load_recipe",
"VideoEncoderConfig",
# Defaults
"camera_encoder_defaults",
# Constants
"VALID_VIDEO_CODECS",
"VIDEO_ENCODER_INFO_KEYS",
]
src/lerobot/configs/dataset.py
+6 -5
View File
@@ -14,10 +14,12 @@
"""Shared dataset recording configuration used by both ``lerobot-record`` and ``lerobot-rollout``."""
from dataclasses import dataclass
from dataclasses import dataclass, field
from datetime import datetime
from pathlib import Path
from .video import VideoEncoderConfig, camera_encoder_defaults
@dataclass
class DatasetRecordConfig:
@@ -55,10 +57,9 @@ class DatasetRecordConfig:
# Number of episodes to record before batch encoding videos
# Set to 1 for immediate encoding (default behavior), or higher for batched encoding
video_encoding_batch_size: int = 1
# Video codec for encoding videos. Options: 'h264', 'hevc', 'libsvtav1', 'auto',
# or hardware-specific: 'h264_videotoolbox', 'h264_nvenc', 'h264_vaapi', 'h264_qsv'.
# Use 'auto' to auto-detect the best available hardware encoder.
vcodec: str = "libsvtav1"
# Video encoder settings for camera MP4s (codec, quality, GOP, etc.). Tuned via CLI nested keys,
# e.g. ``--dataset.camera_encoder.vcodec=h264`` (see ``VideoEncoderConfig``).
camera_encoder: VideoEncoderConfig = field(default_factory=camera_encoder_defaults)
# Enable streaming video encoding: encode frames in real-time during capture instead
# of writing PNG images first. Makes save_episode() near-instant. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding
streaming_encoding: bool = False
src/lerobot/configs/default.py
+2 -2
View File
@@ -17,7 +17,7 @@
from dataclasses import dataclass, field
from lerobot.transforms import ImageTransformsConfig
from lerobot.utils.import_utils import get_safe_default_codec
from lerobot.utils.import_utils import get_safe_default_video_backend
@dataclass
@@ -34,7 +34,7 @@ class DatasetConfig:
image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)
revision: str | None = None
use_imagenet_stats: bool = True
video_backend: str = field(default_factory=get_safe_default_codec)
video_backend: str = field(default_factory=get_safe_default_video_backend)
# When True, video frames are returned as uint8 tensors (0-255) instead of float32 (0.0-1.0).
# This reduces memory and speeds up DataLoader IPC. The training pipeline handles the conversion.
return_uint8: bool = False
src/lerobot/configs/eval.py
+1 -1
View File
@@ -18,8 +18,8 @@ from logging import getLogger
from pathlib import Path
from lerobot import envs, policies # noqa: F401
from lerobot.configs import parser
from . import parser
from .default import EvalConfig
from .policies import PreTrainedConfig
src/lerobot/configs/recipe.py
+206
View File
@@ -0,0 +1,206 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations
import re
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Literal, get_args
MessageRole = Literal["user", "assistant", "system", "tool"]
MessageStream = Literal["high_level", "low_level"]
DEFAULT_BINDINGS = {
"subtask": "active_at(t, style=subtask)",
"memory": "active_at(t, style=memory)",
"plan": "active_at(t, style=plan)",
"speech": "emitted_at(t, role=assistant, tool_name=say)",
"interjection": "emitted_at(t, style=interjection)",
"vqa": "emitted_at(t, style=vqa, role=assistant)",
"vqa_query": "emitted_at(t, style=vqa, role=user)",
}
PLACEHOLDER_RE = re.compile(r"\$\{([A-Za-z_][A-Za-z0-9_]*)\}")
"""``${name}`` placeholder pattern used by both recipe binding-reference
discovery (here) and rendered-message substitution (in ``language_render``)."""
_VALID_ROLES = frozenset(get_args(MessageRole))
_VALID_STREAMS = frozenset(get_args(MessageStream))
@dataclass
class MessageTurn:
"""A single chat-style turn in a recipe template.
``content`` may be a plain string, a list of HF-style multimodal blocks, or
``None`` when ``tool_calls_from`` supplies tool-call payloads instead.
``stream`` tags the turn for downstream filtering, ``target`` flags it as a
training target, and ``if_present`` skips the turn when the named binding
resolves to ``None``.
"""
role: MessageRole
content: str | list[dict[str, Any]] | None = None
stream: MessageStream | None = None
target: bool = False
if_present: str | None = None
tool_calls_from: str | None = None
def __post_init__(self) -> None:
"""Validate role, stream, and content after dataclass construction."""
if self.role not in _VALID_ROLES:
raise ValueError(f"Unsupported message role: {self.role!r}")
# ``stream`` is typed Optional only so the dataclass can keep its
# field ordering, but recipes must always tag every turn with a
# stream — the renderer's ``_validate_rendered`` would reject
# ``None`` later on. Fail at construction so the bad recipe is
# caught at YAML load time rather than at the first sample.
if self.stream is None:
raise ValueError(
f"MessageTurn(role={self.role!r}) is missing a stream — "
f"every turn must declare one of {sorted(_VALID_STREAMS)}."
)
if self.stream not in _VALID_STREAMS:
raise ValueError(f"Unsupported message stream: {self.stream!r}")
if self.content is None and self.tool_calls_from is None:
raise ValueError("MessageTurn.content is required unless tool_calls_from is set.")
if self.content is not None and not isinstance(self.content, (str, list)):
raise TypeError("MessageTurn.content must be a string, a list of HF-style blocks, or None.")
if isinstance(self.content, list):
for block in self.content:
if not isinstance(block, dict) or "type" not in block:
raise ValueError(
"Multimodal content blocks must be HF-style dictionaries with a type key."
)
@classmethod
def from_dict(cls, data: dict[str, Any]) -> MessageTurn:
"""Construct a :class:`MessageTurn` from a plain dictionary."""
return cls(**data)
@dataclass
class TrainingRecipe:
"""A recipe describing how to render training samples from language rows.
A recipe is either a *message recipe* (``messages`` plus optional
``bindings``) or a *blend recipe* (``blend`` mapping names to weighted
sub-recipes). ``weight`` is only meaningful inside a blend.
"""
messages: list[MessageTurn] | None = None
bindings: dict[str, str] | None = None
blend: dict[str, TrainingRecipe] | None = None
weight: float | None = None
def __post_init__(self) -> None:
"""Validate that exactly one of ``messages`` or ``blend`` is set."""
if self.messages is not None and self.blend is not None:
raise ValueError("TrainingRecipe must set only one of messages or blend.")
if self.messages is None and self.blend is None:
raise ValueError("TrainingRecipe must set one of messages or blend.")
if self.messages is not None:
self._validate_message_recipe()
if self.blend is not None:
self._validate_blend_recipe()
@classmethod
def from_dict(cls, data: dict[str, Any]) -> TrainingRecipe:
"""Construct a :class:`TrainingRecipe` from a nested dictionary."""
data = dict(data)
if data.get("messages") is not None:
data["messages"] = [
turn if isinstance(turn, MessageTurn) else MessageTurn.from_dict(turn)
for turn in data["messages"]
]
if data.get("blend") is not None:
data["blend"] = {
name: recipe if isinstance(recipe, TrainingRecipe) else cls.from_dict(recipe)
for name, recipe in data["blend"].items()
}
return cls(**data)
@classmethod
def from_yaml(cls, path: str | Path) -> TrainingRecipe:
"""Load a :class:`TrainingRecipe` from a YAML file at ``path``."""
import yaml # type: ignore[import-untyped]
with open(path) as f:
data = yaml.safe_load(f)
if not isinstance(data, dict):
raise ValueError(f"Recipe YAML must contain a mapping at the top level: {path}")
return cls.from_dict(data)
def _validate_message_recipe(self) -> None:
"""Ensure every templated binding is known and at least one turn is a target."""
assert self.messages is not None
known_bindings = set(DEFAULT_BINDINGS) | set(self.bindings or {}) | {"task"}
for turn in self.messages:
missing = self._referenced_bindings(turn) - known_bindings
if missing:
raise ValueError(f"MessageTurn references unknown binding(s): {sorted(missing)}")
if not any(turn.target for turn in self.messages):
raise ValueError("Message recipes must contain at least one target turn.")
def _validate_blend_recipe(self) -> None:
"""Ensure each blend component is a non-empty, weighted message recipe."""
assert self.blend is not None
if not self.blend:
raise ValueError("Blend recipes must contain at least one component.")
for name, recipe in self.blend.items():
if recipe.blend is not None:
raise ValueError(f"Blend component {name!r} cannot itself define a blend.")
if recipe.messages is None:
raise ValueError(f"Blend component {name!r} must define messages.")
if recipe.weight is None:
raise ValueError(f"Blend component {name!r} must define weight.")
if recipe.weight <= 0:
raise ValueError(f"Blend component {name!r} must have a positive weight.")
def _referenced_bindings(self, turn: MessageTurn) -> set[str]:
"""Return the binding names that ``turn`` references via placeholders or attributes."""
names: set[str] = set()
if turn.if_present is not None:
names.add(turn.if_present)
if turn.tool_calls_from is not None:
names.add(turn.tool_calls_from)
names.update(_placeholders_in_content(turn.content))
return names
def _placeholders_in_content(content: str | list[dict[str, Any]] | None) -> set[str]:
"""Return the set of ``${name}`` placeholders found anywhere in ``content``."""
if content is None:
return set()
if isinstance(content, str):
return set(PLACEHOLDER_RE.findall(content))
names: set[str] = set()
for block in content:
for value in block.values():
if isinstance(value, str):
names.update(PLACEHOLDER_RE.findall(value))
return names
def load_recipe(path: str | Path) -> TrainingRecipe:
"""Load a :class:`TrainingRecipe` from a YAML file at ``path``."""
return TrainingRecipe.from_yaml(path)
src/lerobot/configs/rewards.py
+5 -4
View File
@@ -27,12 +27,13 @@ from huggingface_hub import hf_hub_download
from huggingface_hub.constants import CONFIG_NAME
from huggingface_hub.errors import HfHubHTTPError
from lerobot.configs.types import PolicyFeature
from lerobot.optim.optimizers import OptimizerConfig
from lerobot.optim.schedulers import LRSchedulerConfig
from lerobot.utils.device_utils import auto_select_torch_device, is_torch_device_available
from lerobot.utils.hub import HubMixin
from .types import PolicyFeature
T = TypeVar("T", bound="RewardModelConfig")
logger = logging.getLogger(__name__)
@@ -89,9 +90,9 @@ class RewardModelConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
def reward_delta_indices(self) -> list | None: # type: ignore[type-arg]
return None
@abc.abstractmethod
def get_optimizer_preset(self) -> OptimizerConfig:
raise NotImplementedError
def get_optimizer_preset(self) -> OptimizerConfig | None:
"""Default optimizer for this reward model, or ``None`` for zero-shot models."""
return None
def get_scheduler_preset(self) -> LRSchedulerConfig | None:
return None
src/lerobot/configs/train.py
+1 -1
View File
@@ -25,11 +25,11 @@ from huggingface_hub import hf_hub_download
from huggingface_hub.errors import HfHubHTTPError
from lerobot import envs
from lerobot.configs import parser
from lerobot.optim import LRSchedulerConfig, OptimizerConfig
from lerobot.utils.hub import HubMixin
from lerobot.utils.sample_weighting import SampleWeightingConfig
from . import parser
from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
from .policies import PreTrainedConfig
from .rewards import RewardModelConfig
src/lerobot/configs/video.py
+235
View File
@@ -0,0 +1,235 @@
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Note: We subclass str so that serialization is straightforward
# https://stackoverflow.com/questions/24481852/serialising-an-enum-member-to-json
"""Video encoder configurations."""
from __future__ import annotations
import logging
from dataclasses import dataclass, field
from typing import Any
from lerobot.utils.import_utils import require_package
logger = logging.getLogger(__name__)
# List of hardware encoders to probe for auto-selection. Availability depends on the platform and the chosen video backend.
# Determines the order of preference for auto-selection when vcodec="auto" is used.
HW_VIDEO_CODECS = [
"h264_videotoolbox", # macOS
"hevc_videotoolbox", # macOS
"h264_nvenc", # NVIDIA GPU
"hevc_nvenc", # NVIDIA GPU
"h264_vaapi", # Linux Intel/AMD
"h264_qsv", # Intel Quick Sync
]
VALID_VIDEO_CODECS: frozenset[str] = frozenset({"h264", "hevc", "libsvtav1", "auto", *HW_VIDEO_CODECS})
# Aliases for legacy video codec names.
VIDEO_CODECS_ALIASES: dict[str, str] = {"av1": "libsvtav1"}
LIBSVTAV1_DEFAULT_PRESET: int = 12
# Keys persisted under ``features[*]["info"]`` as ``video.<name>`` (from :class:`VideoEncoderConfig`).
# ``vcodec``` and ``pix_fmt`` are derived from the video stream directly.
VIDEO_ENCODER_INFO_FIELD_NAMES: frozenset[str] = frozenset(
{"g", "crf", "preset", "fast_decode", "extra_options", "video_backend"}
)
VIDEO_ENCODER_INFO_KEYS: frozenset[str] = frozenset(
f"video.{name}" for name in VIDEO_ENCODER_INFO_FIELD_NAMES
)
@dataclass
class VideoEncoderConfig:
"""Video encoder configuration.
Attributes:
vcodec: Video encoder name. ``"auto"`` is resolved during
construction (HW encoder if available, else ``libsvtav1``).
pix_fmt: Pixel format (e.g. ``"yuv420p"``).
g: GOP size (keyframe interval).
crf: Quality level — mapped to the native quality parameter of the
codec (``crf`` for software, ``qp`` for NVENC/VAAPI,
``q:v`` for VideoToolbox, ``global_quality`` for QSV).
preset: Speed/quality preset. Accepted type is per-codec.
fast_decode: Fast-decode tuning. For ``libsvtav1`` this is a level (0-2)
embedded in ``svtav1-params``. For ``h264`` and ``hevc`` non-zero values
set ``tune=fastdecode``. Ignored for other codecs.
video_backend: Python to be used for encoding. Only ``"pyav"``
is currently supported.
extra_options: Free-form dictionary of additional video encoder options
(e.g. ``{"tune": "film", "profile:v": "high", "bf": 2}``).
"""
vcodec: str = "libsvtav1" # TODO(CarolinePascal): rename to codec ?
pix_fmt: str = "yuv420p"
g: int | None = 2
crf: int | float | None = 30
preset: int | str | None = None
fast_decode: int = 0
# TODO(CarolinePascal): add torchcodec support + find a way to unify the
# two backends (encoding and decoding).
video_backend: str = "pyav"
extra_options: dict[str, Any] = field(default_factory=dict)
def __post_init__(self) -> None:
self.resolve_vcodec()
# Empty-constructor ergonomics: ``VideoEncoderConfig()`` must "just work".
if self.preset is None and self.vcodec == "libsvtav1":
self.preset = LIBSVTAV1_DEFAULT_PRESET
self.validate()
@classmethod
def from_video_info(cls, video_info: dict | None) -> VideoEncoderConfig:
"""Reconstruct a :class:`VideoEncoderConfig` from a video feature's ``info`` block.
Missing or ``None`` values fall back to the class defaults.
"""
video_info = video_info or {}
kwargs: dict[str, Any] = {}
for src_key, dst_field in (("video.codec", "vcodec"), ("video.pix_fmt", "pix_fmt")):
value = video_info.get(src_key)
if value is not None:
kwargs[dst_field] = value
for field_name in VIDEO_ENCODER_INFO_FIELD_NAMES:
value = video_info.get(f"video.{field_name}")
if value is None:
continue
# Persisted as ``{}`` after merges with disagreeing sources — treat as default.
if field_name == "extra_options" and not value:
continue
kwargs[field_name] = value
return cls(**kwargs)
def detect_available_encoders(self, encoders: list[str] | str) -> list[str]:
"""Return the subset of available encoders based on the specified video backend.
Args:
encoders: List of encoder names to detect. If a string, it is converted to a list.
Returns:
List of available encoder names. If the video backend is not "pyav", returns an empty list.
"""
if self.video_backend == "pyav":
require_package("av", extra="dataset")
from lerobot.datasets import detect_available_encoders_pyav
return detect_available_encoders_pyav(encoders)
return []
def validate(self) -> None:
"""Validate the video encoder configuration."""
if self.video_backend == "pyav":
require_package("av", extra="dataset")
from lerobot.datasets import check_video_encoder_parameters_pyav
check_video_encoder_parameters_pyav(self.vcodec, self.pix_fmt, self.get_codec_options())
def resolve_vcodec(self) -> None:
"""Check ``vcodec`` and, when it is ``"auto"``, pick a concrete encoder.
For ``"auto"``, the first hardware encoder in the preference list that is available is chosen; if none are available, ``libsvtav1`` is used. If the
resolved codec (explicit or after auto-selection) is not available, raises ``ValueError``.
Stream-derived canonical codec names listed in :data:`VIDEO_CODECS_ALIASES` are
rewritten to their corresponding encoder name (e.g. ``"av1"`` → ``"libsvtav1"``).
"""
self.vcodec = VIDEO_CODECS_ALIASES.get(self.vcodec, self.vcodec)
if self.vcodec not in VALID_VIDEO_CODECS:
raise ValueError(f"Invalid vcodec '{self.vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
if self.vcodec == "auto":
available = self.detect_available_encoders(HW_VIDEO_CODECS)
for encoder in HW_VIDEO_CODECS:
if encoder in available:
logger.info(f"Auto-selected video codec: {encoder}")
self.vcodec = encoder
return
logger.warning("No hardware encoder available, falling back to software encoder 'libsvtav1'")
self.vcodec = "libsvtav1"
if self.detect_available_encoders(self.vcodec):
logger.info(f"Using video codec: {self.vcodec}")
return
raise ValueError(f"Unsupported video codec: {self.vcodec} with video backend {self.video_backend}")
def get_codec_options(
self, encoder_threads: int | None = None, as_strings: bool = False
) -> dict[str, Any]:
"""Translate the tuning fields to codec-specific options.
``VideoEncoderConfig.extra_options`` are merged last but never override a structured field.
Args:
encoder_threads: Number of encoder threads set globally for all VideoEncoderConfigs.
For libsvtav1, this is mapped to ``lp`` via ``svtav1-params``.
For h264/hevc, this is mapped to ``threads``.
Hardware encoders ignore this parameter.
as_strings: If ``True``, casts values to strings.
"""
opts: dict[str, Any] = {}
def set_if(key: str, value: Any) -> None:
if value is not None:
opts[key] = value if not as_strings else str(value)
# GOP size is not a codec-specific option, so it is always set.
set_if("g", self.g)
if self.vcodec == "libsvtav1":
set_if("crf", self.crf)
set_if("preset", self.preset)
svtav1_parts: list[str] = []
if self.fast_decode is not None:
svtav1_parts.append(f"fast-decode={max(0, min(2, self.fast_decode))}")
if encoder_threads is not None:
svtav1_parts.append(f"lp={encoder_threads}")
if svtav1_parts:
opts["svtav1-params"] = ":".join(svtav1_parts)
elif self.vcodec in ("h264", "hevc"):
set_if("crf", self.crf)
set_if("preset", self.preset)
if self.fast_decode:
opts["tune"] = "fastdecode"
set_if("threads", encoder_threads)
elif self.vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
if self.crf is not None:
opts["q:v"] = max(1, min(100, 100 - self.crf * 2))
elif self.vcodec in ("h264_nvenc", "hevc_nvenc"):
opts["rc"] = 0
set_if("qp", self.crf)
set_if("preset", self.preset)
elif self.vcodec == "h264_vaapi":
set_if("qp", self.crf)
elif self.vcodec == "h264_qsv":
set_if("global_quality", self.crf)
set_if("preset", self.preset)
else:
set_if("crf", self.crf)
set_if("preset", self.preset)
# Extra options are merged last but never override structured fields (values are kept as given).
for k, v in self.extra_options.items():
if k not in opts:
set_if(k, v)
return opts
def camera_encoder_defaults() -> VideoEncoderConfig:
"""Return a :class:`VideoEncoderConfig` with RGB-camera defaults."""
return VideoEncoderConfig()
src/lerobot/datasets/__init__.py
+19
View File
@@ -31,15 +31,25 @@ from .dataset_tools import (
modify_features,
modify_tasks,
recompute_stats,
reencode_dataset,
remove_feature,
split_dataset,
)
from .factory import make_dataset, resolve_delta_timestamps
from .image_writer import safe_stop_image_writer
from .io_utils import load_episodes, write_stats
from .language import (
EVENT_ONLY_STYLES,
LANGUAGE_EVENTS,
LANGUAGE_PERSISTENT,
PERSISTENT_STYLES,
STYLE_REGISTRY,
column_for_style,
)
from .lerobot_dataset import LeRobotDataset
from .multi_dataset import MultiLeRobotDataset
from .pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
from .pyav_utils import check_video_encoder_parameters_pyav, detect_available_encoders_pyav
from .sampler import EpisodeAwareSampler
from .streaming_dataset import StreamingLeRobotDataset
from .utils import DEFAULT_EPISODES_PATH, create_lerobot_dataset_card
@@ -53,12 +63,19 @@ __all__ = [
"CODEBASE_VERSION",
"DEFAULT_EPISODES_PATH",
"DEFAULT_QUANTILES",
"EVENT_ONLY_STYLES",
"EpisodeAwareSampler",
"LANGUAGE_EVENTS",
"LANGUAGE_PERSISTENT",
"LeRobotDataset",
"LeRobotDatasetMetadata",
"MultiLeRobotDataset",
"PERSISTENT_STYLES",
"STYLE_REGISTRY",
"StreamingLeRobotDataset",
"VideoEncodingManager",
"check_video_encoder_parameters_pyav",
"detect_available_encoders_pyav",
"add_features",
"aggregate_datasets",
"aggregate_pipeline_dataset_features",
@@ -66,6 +83,7 @@ __all__ = [
"convert_image_to_video_dataset",
"create_initial_features",
"create_lerobot_dataset_card",
"column_for_style",
"delete_episodes",
"get_feature_stats",
"load_episodes",
@@ -74,6 +92,7 @@ __all__ = [
"modify_features",
"modify_tasks",
"recompute_stats",
"reencode_dataset",
"remove_feature",
"resolve_delta_timestamps",
"safe_stop_image_writer",
src/lerobot/datasets/aggregate.py
+52 -4
View File
@@ -15,6 +15,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import copy
import logging
import shutil
from pathlib import Path
@@ -23,9 +24,11 @@ import datasets
import pandas as pd
import tqdm
from lerobot.configs import VIDEO_ENCODER_INFO_KEYS
from .compute_stats import aggregate_stats
from .dataset_metadata import LeRobotDatasetMetadata
from .feature_utils import get_hf_features_from_features
from .feature_utils import features_equal_for_merge, get_hf_features_from_features
from .io_utils import (
get_file_size_in_mb,
get_parquet_file_size_in_mb,
@@ -46,11 +49,54 @@ from .utils import (
from .video_utils import concatenate_video_files, get_video_duration_in_s
def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMetadata]) -> dict[str, dict]:
"""Create a merged video feature info dictionary for aggregation. The video encoder info is merged field-by-field: each key is kept only when every source agrees; otherwise that key is set to ``null`` (or ``{}`` for ``video.extra_options``) and a warning is logged.
Args:
all_metadata: List of LeRobotDatasetMetadata objects to merge.
Returns:
dict: A dictionary of merged video feature info.
"""
merged_info = copy.deepcopy(all_metadata[0].features)
video_keys = [k for k in merged_info if merged_info[k].get("dtype") == "video"]
for vk in video_keys:
video_infos = [m.features.get(vk, {}).get("info") or {} for m in all_metadata]
base_video_info = video_infos[0]
merged_encoder_info: dict = {}
fallback_keys: list[str] = []
for info_key in VIDEO_ENCODER_INFO_KEYS:
values = [info.get(info_key, None) for info in video_infos]
first_value = values[0]
all_match = all(v == first_value for v in values[1:])
if all_match:
merged_encoder_info[info_key] = first_value
else:
fallback_keys.append(info_key)
merged_encoder_info[info_key] = {} if info_key == "video.extra_options" else None
if fallback_keys:
logging.warning(
f"Merging heterogeneous or incomplete video encoder metadata for feature {vk}. "
f"Setting these keys to null: {fallback_keys}.",
)
merged_info[vk]["info"] = {**base_video_info, **merged_encoder_info}
# TODO(CarolinePascal): make this variable once we have support for other video backends.
merged_info[vk]["info"]["video.video_backend"] = "pyav"
return merged_info
def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
"""Validates that all dataset metadata have consistent properties.
Ensures all datasets have the same fps, robot_type, and features to guarantee
compatibility when aggregating them into a single dataset.
Video encoder info is not considered for validation but is merged during aggregation in ``merge_video_feature_info_for_aggregate``.
Args:
all_metadata: List of LeRobotDatasetMetadata objects to validate.
@@ -74,7 +120,7 @@ def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
raise ValueError(
f"Same robot_type is expected, but got robot_type={meta.robot_type} instead of {robot_type}."
)
if features != meta.features:
if not features_equal_for_merge(features, meta.features):
raise ValueError(
f"Same features is expected, but got features={meta.features} instead of {features}."
)
@@ -274,7 +320,8 @@ def aggregate_datasets(
LeRobotDatasetMetadata(repo_id, root=root) for repo_id, root in zip(repo_ids, roots, strict=False)
]
)
fps, robot_type, features = validate_all_metadata(all_metadata)
fps, robot_type, _ = validate_all_metadata(all_metadata)
features = merge_video_feature_info_for_aggregate(all_metadata)
video_keys = [key for key in features if features[key]["dtype"] == "video"]
dst_meta = LeRobotDatasetMetadata.create(
@@ -332,7 +379,6 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
videos_idx: Dictionary tracking video chunk and file indices.
video_files_size_in_mb: Maximum size for video files in MB (defaults to DEFAULT_VIDEO_FILE_SIZE_IN_MB)
chunk_size: Maximum number of files per chunk (defaults to DEFAULT_CHUNK_SIZE)
Returns:
dict: Updated videos_idx with current chunk and file indices.
"""
@@ -414,9 +460,11 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
current_dst_duration = dst_file_durations.get(dst_key, 0)
videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_dst_duration
videos_idx[key]["src_to_dst"][(src_chunk_idx, src_file_idx)] = dst_key
# TODO(CarolinePascal): Move the check before the loop to avoid failing in the middle + add possibility to re-encode the video if the check fails
concatenate_video_files(
[dst_path, src_path],
dst_path,
compatibility_check=True,
)
# Update duration of this destination file
dst_file_durations[dst_key] = current_dst_duration + src_duration
src/lerobot/datasets/compute_stats.py
+1 -1
View File
@@ -512,7 +512,7 @@ def compute_episode_stats(
ep_stats = {}
for key, data in episode_data.items():
if features[key]["dtype"] == "string":
if features[key]["dtype"] in {"string", "language"}:
continue
if features[key]["dtype"] in ["image", "video"]:
src/lerobot/datasets/dataset_metadata.py
+61 -6
View File
@@ -24,6 +24,7 @@ import pyarrow as pa
import pyarrow.parquet as pq
from huggingface_hub import snapshot_download
from lerobot.configs import VideoEncoderConfig
from lerobot.utils.constants import DEFAULT_FEATURES, HF_LEROBOT_HOME, HF_LEROBOT_HUB_CACHE
from lerobot.utils.feature_utils import _validate_feature_names
from lerobot.utils.utils import flatten_dict
@@ -35,12 +36,12 @@ from .io_utils import (
load_episodes,
load_info,
load_stats,
load_subtasks,
load_tasks,
write_info,
write_stats,
write_tasks,
)
from .language import DEFAULT_TOOLS, LANGUAGE_COLUMNS
from .utils import (
DEFAULT_EPISODES_PATH,
check_version_compatibility,
@@ -176,7 +177,6 @@ class LeRobotDatasetMetadata:
self.info = load_info(self.root)
check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
self.tasks = load_tasks(self.root)
self.subtasks = load_subtasks(self.root)
self.episodes = load_episodes(self.root)
self.stats = load_stats(self.root)
@@ -342,6 +342,49 @@ class LeRobotDatasetMetadata:
"""Keys to access visual modalities (regardless of their storage method)."""
return [key for key, ft in self.features.items() if ft["dtype"] in ["video", "image"]]
@property
def has_language_columns(self) -> bool:
"""Return ``True`` if the dataset declares any language column.
Used to gate language-aware code paths (collate, render step) so
unannotated datasets keep PyTorch's default collate behavior.
"""
return any(col in self.features for col in LANGUAGE_COLUMNS)
@property
def tools(self) -> list[dict]:
"""OpenAI-style tool schemas declared by this dataset.
Read from ``meta/info.json["tools"]``. Returns a copy, so callers
can mutate the result safely. Falls back to
:data:`lerobot.datasets.language.DEFAULT_TOOLS` (the canonical
``say`` schema) when the dataset doesn't declare any — that way
unannotated datasets and chat-template consumers
(``apply_chat_template(messages, tools=meta.tools)``) keep
working out of the box.
Implementations live under :mod:`lerobot.tools` (one file per
tool); see ``docs/source/tools.mdx`` for the authoring guide.
"""
declared = self.info.tools
if declared:
return [dict(t) for t in declared]
return [dict(t) for t in DEFAULT_TOOLS]
@tools.setter
def tools(self, value: list[dict] | None) -> None:
"""Persist a tool catalog to ``meta/info.json`` and reload metadata.
Writes ``value`` into the on-disk ``info.json`` (or clears the
``tools`` key when ``value`` is ``None`` or empty), then reloads
``self.info`` so the in-memory metadata matches what's on disk.
Saves callers from hand-editing ``info.json`` and re-instantiating
the metadata object.
"""
self.info.tools = [dict(t) for t in value] if value else None
write_info(self.info, self.root)
self.info = load_info(self.root)
@property
def names(self) -> dict[str, list | dict]:
"""Names of the various dimensions of vector modalities."""
@@ -534,10 +577,23 @@ class LeRobotDatasetMetadata:
self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats is not None else episode_stats
write_stats(self.stats, self.root)
def update_video_info(self, video_key: str | None = None) -> None:
"""
def update_video_info(
self,
video_key: str | None = None,
camera_encoder: VideoEncoderConfig | None = None,
) -> None:
"""Populate per-feature video info in ``info.json``.
Warning: this function writes info from first episode videos, implicitly assuming that all videos have
been encoded the same way. Also, this means it assumes the first episode exists.
Args:
video_key: If provided, only update this video key. Otherwise update
all video keys in the dataset.
camera_encoder: Encoder configuration used to produce the
videos. When provided, its fields are recorded as
``video.<field>`` entries alongside the stream-derived
``video.*`` entries (see :func:`get_video_info`).
"""
if video_key is not None and video_key not in self.video_keys:
raise ValueError(f"Video key {video_key} not found in dataset")
@@ -546,7 +602,7 @@ class LeRobotDatasetMetadata:
for key in video_keys:
if not self.features[key].get("info", None):
video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
self.info.features[key]["info"] = get_video_info(video_path)
self.info.features[key]["info"] = get_video_info(video_path, camera_encoder=camera_encoder)
def update_chunk_settings(
self,
@@ -657,7 +713,6 @@ class LeRobotDatasetMetadata:
_validate_feature_names(features)
obj.tasks = None
obj.subtasks = None
obj.episodes = None
obj.stats = None
obj.info = create_empty_dataset_info(
src/lerobot/datasets/dataset_reader.py
-5
View File
@@ -295,9 +295,4 @@ class DatasetReader:
task_idx = item["task_index"].item()
item["task"] = self._meta.tasks.iloc[task_idx].name
# add subtask information if available
if "subtask_index" in self._meta.features and self._meta.subtasks is not None:
subtask_idx = item["subtask_index"].item()
item["subtask"] = self._meta.subtasks.iloc[subtask_idx].name
return item
src/lerobot/datasets/dataset_tools.py
+128 -50
View File
@@ -26,7 +26,7 @@ This module provides utilities for:
import logging
import shutil
from collections.abc import Callable
from concurrent.futures import ThreadPoolExecutor, as_completed
from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor, as_completed
from pathlib import Path
import datasets
@@ -36,6 +36,7 @@ import pyarrow.parquet as pq
import torch
from tqdm import tqdm
from lerobot.configs import VideoEncoderConfig, camera_encoder_defaults
from lerobot.utils.constants import ACTION, HF_LEROBOT_HOME, OBS_IMAGE, OBS_STATE
from lerobot.utils.utils import flatten_dict
@@ -60,9 +61,14 @@ from .utils import (
DEFAULT_DATA_FILE_SIZE_IN_MB,
DEFAULT_DATA_PATH,
DEFAULT_EPISODES_PATH,
VIDEO_DIR,
update_chunk_file_indices,
)
from .video_utils import encode_video_frames, get_video_info
from .video_utils import (
encode_video_frames,
get_video_info,
reencode_video,
)
def _load_episode_with_stats(src_dataset: LeRobotDataset, episode_idx: int) -> dict:
@@ -95,6 +101,11 @@ def delete_episodes(
) -> LeRobotDataset:
"""Delete episodes from a LeRobotDataset and create a new dataset.
Video segments that need re-encoding (because the source file mixes kept and
deleted episodes) are re-encoded with the source dataset's existing encoder
settings — read back from ``meta/info.json`` — so the output dataset stays
consistent with its own metadata.
Args:
dataset: The source LeRobotDataset.
episode_indices: List of episode indices to delete.
@@ -157,6 +168,11 @@ def split_dataset(
) -> dict[str, LeRobotDataset]:
"""Split a LeRobotDataset into multiple smaller datasets.
Video segments that need re-encoding (because the source file mixes episodes
that fall into different splits) are re-encoded with the source dataset's
existing encoder settings — read back from ``meta/info.json`` — so each
output split stays consistent with its own metadata.
Args:
dataset: The source LeRobotDataset to split.
splits: Either a dict mapping split names to episode indices, or a dict mapping
@@ -578,8 +594,7 @@ def _keep_episodes_from_video_with_av(
output_path: Path,
episodes_to_keep: list[tuple[int, int]],
fps: float,
vcodec: str = "libsvtav1",
pix_fmt: str = "yuv420p",
camera_encoder: VideoEncoderConfig,
) -> None:
"""Keep only specified episodes from a video file using PyAV.
@@ -593,8 +608,7 @@ def _keep_episodes_from_video_with_av(
Ranges are half-open intervals: [start_frame, end_frame), where start_frame
is inclusive and end_frame is exclusive.
fps: Frame rate of the video.
vcodec: Video codec to use for encoding.
pix_fmt: Pixel format for output video.
camera_encoder: Video encoder settings used to re-encode the kept frames.
"""
from fractions import Fraction
@@ -619,12 +633,13 @@ def _keep_episodes_from_video_with_av(
# Convert fps to Fraction for PyAV compatibility.
fps_fraction = Fraction(fps).limit_denominator(1000)
v_out = out.add_stream(vcodec, rate=fps_fraction)
codec_options = camera_encoder.get_codec_options(as_strings=True)
v_out = out.add_stream(camera_encoder.vcodec, rate=fps_fraction, options=codec_options)
# PyAV type stubs don't distinguish video streams from audio/subtitle streams.
v_out.width = v_in.codec_context.width
v_out.height = v_in.codec_context.height
v_out.pix_fmt = pix_fmt
v_out.pix_fmt = camera_encoder.pix_fmt
# Set time_base to match the frame rate for proper timestamp handling.
v_out.time_base = Fraction(1, int(fps))
@@ -687,14 +702,14 @@ def _copy_and_reindex_videos(
src_dataset: LeRobotDataset,
dst_meta: LeRobotDatasetMetadata,
episode_mapping: dict[int, int],
vcodec: str = "libsvtav1",
pix_fmt: str = "yuv420p",
) -> dict[int, dict]:
"""Copy and filter video files, only re-encoding files with deleted episodes.
For video files that only contain kept episodes, we copy them directly.
For files with mixed kept/deleted episodes, we use PyAV filters to efficiently
re-encode only the desired segments.
re-encode only the desired segments. The encoder used for re-encoding is
derived per video key from the source dataset's ``meta/info.json`` so the
destination metadata keeps describing the videos accurately.
Args:
src_dataset: Source dataset to copy from
@@ -711,6 +726,9 @@ def _copy_and_reindex_videos(
for video_key in src_dataset.meta.video_keys:
logging.info(f"Processing videos for {video_key}")
camera_encoder = VideoEncoderConfig.from_video_info(
src_dataset.meta.info.features.get(video_key, {}).get("info")
)
if dst_meta.video_path is None:
raise ValueError("Destination metadata has no video_path defined")
@@ -792,8 +810,7 @@ def _copy_and_reindex_videos(
dst_video_path,
episodes_to_keep_ranges,
src_dataset.meta.fps,
vcodec,
pix_fmt,
camera_encoder,
)
cumulative_ts = 0.0
@@ -1264,11 +1281,7 @@ def _estimate_frame_size_via_calibration(
episode_indices: list[int],
temp_dir: Path,
fps: int,
vcodec: str,
pix_fmt: str,
g: int,
crf: int,
fast_decode: int,
camera_encoder: VideoEncoderConfig,
num_calibration_frames: int = 30,
) -> float:
"""Estimate MB per frame by encoding a small calibration sample.
@@ -1282,11 +1295,7 @@ def _estimate_frame_size_via_calibration(
episode_indices: List of episode indices being processed.
temp_dir: Temporary directory for calibration files.
fps: Frames per second for video encoding.
vcodec: Video codec (libsvtav1, h264, hevc).
pix_fmt: Pixel format (yuv420p, etc.).
g: GOP size (group of pictures).
crf: Constant Rate Factor (quality).
fast_decode: Fast decode tuning parameter.
camera_encoder: Video encoder settings used for calibration encoding.
num_calibration_frames: Number of frames to use for calibration (default: 30).
Returns:
@@ -1322,11 +1331,7 @@ def _estimate_frame_size_via_calibration(
imgs_dir=calibration_dir,
video_path=calibration_video_path,
fps=fps,
vcodec=vcodec,
pix_fmt=pix_fmt,
g=g,
crf=crf,
fast_decode=fast_decode,
camera_encoder=camera_encoder,
overwrite=True,
)
@@ -1644,11 +1649,7 @@ def convert_image_to_video_dataset(
dataset: LeRobotDataset,
output_dir: Path | None = None,
repo_id: str | None = None,
vcodec: str = "libsvtav1",
pix_fmt: str = "yuv420p",
g: int = 2,
crf: int = 30,
fast_decode: int = 0,
camera_encoder: VideoEncoderConfig | None = None,
episode_indices: list[int] | None = None,
num_workers: int = 4,
max_episodes_per_batch: int | None = None,
@@ -1663,11 +1664,8 @@ def convert_image_to_video_dataset(
dataset: The source LeRobot dataset with images
output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
vcodec: Video codec (default: libsvtav1)
pix_fmt: Pixel format (default: yuv420p)
g: Group of pictures size (default: 2)
crf: Constant rate factor (default: 30)
fast_decode: Fast decode tuning (default: 0)
camera_encoder: Video encoder settings
(``None`` uses :func:`~lerobot.configs.camera_encoder_defaults`).
episode_indices: List of episode indices to convert (None = all episodes)
num_workers: Number of threads for parallel processing (default: 4)
max_episodes_per_batch: Maximum episodes per video batch to avoid memory issues (None = no limit)
@@ -1676,6 +1674,9 @@ def convert_image_to_video_dataset(
Returns:
New LeRobotDataset with images encoded as videos
"""
if camera_encoder is None:
camera_encoder = camera_encoder_defaults()
# Check that it's an image dataset
if len(dataset.meta.video_keys) > 0:
raise ValueError(
@@ -1699,7 +1700,10 @@ def convert_image_to_video_dataset(
logging.info(
f"Converting {len(episode_indices)} episodes with {len(img_keys)} cameras from {dataset.repo_id}"
)
logging.info(f"Video codec: {vcodec}, pixel format: {pix_fmt}, GOP: {g}, CRF: {crf}")
logging.info(
f"Video codec: {camera_encoder.vcodec}, pixel format: {camera_encoder.pix_fmt}, "
f"GOP: {camera_encoder.g}, CRF: {camera_encoder.crf}"
)
# Create new features dict, converting image features to video features
new_features = {}
@@ -1769,11 +1773,7 @@ def convert_image_to_video_dataset(
episode_indices=episode_indices,
temp_dir=temp_dir,
fps=fps,
vcodec=vcodec,
pix_fmt=pix_fmt,
g=g,
crf=crf,
fast_decode=fast_decode,
camera_encoder=camera_encoder,
)
logging.info(f"Processing camera: {img_key}")
@@ -1815,11 +1815,7 @@ def convert_image_to_video_dataset(
imgs_dir=imgs_dir,
video_path=video_path,
fps=fps,
vcodec=vcodec,
pix_fmt=pix_fmt,
g=g,
crf=crf,
fast_decode=fast_decode,
camera_encoder=camera_encoder,
overwrite=True,
)
@@ -1865,7 +1861,9 @@ def convert_image_to_video_dataset(
video_path = new_meta.root / new_meta.video_path.format(
video_key=img_key, chunk_index=0, file_index=0
)
new_meta.info.features[img_key]["info"] = get_video_info(video_path)
new_meta.info.features[img_key]["info"] = get_video_info(
video_path, camera_encoder=camera_encoder
)
write_info(new_meta.info, new_meta.root)
@@ -1888,3 +1886,83 @@ def convert_image_to_video_dataset(
# Return new dataset
return LeRobotDataset(repo_id=repo_id, root=output_dir)
def _reencode_video_worker(args: tuple) -> Path:
"""Picklable worker for :func:`reencode_dataset`'s process pool."""
video_path, camera_encoder, encoder_threads = args
reencode_video(
input_video_path=video_path,
output_video_path=video_path,
camera_encoder=camera_encoder,
encoder_threads=encoder_threads,
overwrite=True,
)
return video_path
def reencode_dataset(
dataset: LeRobotDataset,
camera_encoder: VideoEncoderConfig,
encoder_threads: int | None = None,
num_workers: int | None = None,
) -> LeRobotDataset:
"""Re-encode every video in a dataset with a new set of encoding parameters.
Videos are re-encoded in-place and the video information in ``info.json`` is refreshed.
Args:
dataset: An existing :class:`LeRobotDataset` whose videos will be
re-encoded.
camera_encoder: Target encoder configuration applied to every video
file.
encoder_threads: Per-encoder thread count forwarded to
:func:`reencode_video`. ``None`` lets the codec decide.
num_workers: Number of parallel processes. ``None`` or ``0`` means
sequential (no multiprocessing); ``1+`` spawns a
:class:`~concurrent.futures.ProcessPoolExecutor`.
Returns:
The same :class:`LeRobotDataset` instance with its metadata updated
on disk.
"""
meta = dataset.meta
video_paths_list = []
# Only re-encode if the videos are not already encoded with the given video encoding parameters
for video_key in meta.video_keys:
current_info = meta.info.features[video_key].get("info", {})
current_encoder = VideoEncoderConfig.from_video_info(current_info)
if current_encoder != camera_encoder:
video_paths_list.extend((meta.root / VIDEO_DIR / video_key).rglob("*.mp4"))
else:
logging.info(f"{video_key} videos are already encoded with {camera_encoder}. Nothing to do.")
if len(video_paths_list) == 0:
logging.warning("Dataset has no videos to re-encode.")
return dataset
logging.info(f"Re-encoding {len(video_paths_list)} video file(s) with {camera_encoder}")
worker_args = [(vp, camera_encoder, encoder_threads) for vp in video_paths_list]
if num_workers and num_workers > 1:
with ProcessPoolExecutor(max_workers=num_workers) as pool:
futures = [pool.submit(_reencode_video_worker, args) for args in worker_args]
for future in tqdm(
as_completed(futures),
total=len(futures),
desc="Re-encoding videos",
):
future.result()
else:
for args in tqdm(worker_args, desc="Re-encoding videos"):
_reencode_video_worker(args)
# Refresh video info in metadata for every video key.
for vid_key in meta.video_keys:
video_path = meta.root / meta.get_video_file_path(0, vid_key)
meta.info.features[vid_key]["info"] = get_video_info(video_path, camera_encoder=camera_encoder)
write_info(meta.info, meta.root)
logging.info("Dataset metadata updated.")
return dataset
src/lerobot/datasets/dataset_writer.py
+32 -11
View File
@@ -31,6 +31,8 @@ import PIL.Image
import pyarrow.parquet as pq
import torch
from lerobot.configs import VideoEncoderConfig, camera_encoder_defaults
from .compute_stats import compute_episode_stats
from .dataset_metadata import LeRobotDatasetMetadata
from .feature_utils import (
@@ -65,14 +67,19 @@ def _encode_video_worker(
episode_index: int,
root: Path,
fps: int,
vcodec: str = "libsvtav1",
camera_encoder: VideoEncoderConfig | None = None,
encoder_threads: int | None = None,
) -> Path:
temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
img_dir = (root / fpath).parent
encode_video_frames(
img_dir, temp_path, fps, vcodec=vcodec, overwrite=True, encoder_threads=encoder_threads
img_dir,
temp_path,
fps,
camera_encoder=camera_encoder,
encoder_threads=encoder_threads,
overwrite=True,
)
shutil.rmtree(img_dir)
return temp_path
@@ -89,20 +96,22 @@ class DatasetWriter:
self,
meta: LeRobotDatasetMetadata,
root: Path,
vcodec: str,
camera_encoder: VideoEncoderConfig | None,
encoder_threads: int | None,
batch_encoding_size: int,
streaming_encoder: StreamingVideoEncoder | None = None,
initial_frames: int = 0,
):
"""Initialize the writer with metadata, codec, and encoding config.
"""Initialize the writer with metadata, codec, and encoder config.
Args:
meta: Dataset metadata instance (used for feature schema, chunk
settings, and episode persistence).
root: Local dataset root directory.
vcodec: Video codec for encoding (e.g. ``'libsvtav1'``, ``'h264'``).
encoder_threads: Threads per encoder instance. ``None`` for auto.
camera_encoder: Video encoder settings applied to all cameras.
``None`` uses :func:`~lerobot.configs.camera_encoder_defaults`.
encoder_threads: Number of encoder threads (global). ``None``
lets the codec decide.
batch_encoding_size: Number of episodes to accumulate before
batch-encoding videos.
streaming_encoder: Optional pre-built :class:`StreamingVideoEncoder`
@@ -111,7 +120,7 @@ class DatasetWriter:
"""
self._meta = meta
self._root = root
self._vcodec = vcodec
self._camera_encoder = camera_encoder or camera_encoder_defaults()
self._encoder_threads = encoder_threads
self._batch_encoding_size = batch_encoding_size
self._streaming_encoder = streaming_encoder
@@ -241,7 +250,14 @@ class DatasetWriter:
for key, ft in self._meta.features.items():
if key in ["index", "episode_index", "task_index"] or ft["dtype"] in ["image", "video"]:
continue
episode_buffer[key] = np.stack(episode_buffer[key])
stacked_values = np.stack(episode_buffer[key])
# `shape=(1,)` numeric features are serialized as `datasets.Value`, which expects scalars.
# Normalizing to `(N,)` keeps save semantics stable across dependency versions.
if tuple(ft["shape"]) == (1,) and ft["dtype"] != "string":
stacked_values = stacked_values.reshape(episode_length)
episode_buffer[key] = stacked_values
# Wait for image writer to end, so that episode stats over images can be computed
self._wait_image_writer()
@@ -284,7 +300,7 @@ class DatasetWriter:
episode_index,
self._root,
self._meta.fps,
self._vcodec,
self._camera_encoder,
self._encoder_threads,
): video_key
for video_key in self._meta.video_keys
@@ -495,7 +511,7 @@ class DatasetWriter:
# Update video info (only needed when first episode is encoded)
if episode_index == 0:
self._meta.update_video_info(video_key)
self._meta.update_video_info(video_key, camera_encoder=self._camera_encoder)
write_info(self._meta.info, self._meta.root)
metadata = {
@@ -564,7 +580,12 @@ class DatasetWriter:
def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> Path:
"""Use ffmpeg to convert frames stored as png into mp4 videos."""
return _encode_video_worker(
video_key, episode_index, self._root, self._meta.fps, self._vcodec, self._encoder_threads
video_key,
episode_index,
self._root,
self._meta.fps,
self._camera_encoder,
self._encoder_threads,
)
def close_writer(self) -> None:
src/lerobot/datasets/feature_utils.py
+76 -1
View File
@@ -13,15 +13,23 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
from pprint import pformat
import datasets
import numpy as np
from PIL import Image as PILImage
from lerobot.configs import VIDEO_ENCODER_INFO_KEYS
from lerobot.utils.constants import DEFAULT_FEATURES
from lerobot.utils.utils import is_valid_numpy_dtype_string
from .language import (
LANGUAGE_PERSISTENT,
is_language_column,
language_events_column_feature,
language_persistent_column_feature,
)
from .utils import (
DEFAULT_CHUNK_SIZE,
DEFAULT_DATA_FILE_SIZE_IN_MB,
@@ -46,7 +54,13 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
"""
hf_features = {}
for key, ft in features.items():
if ft["dtype"] == "video":
if is_language_column(key):
hf_features[key] = (
language_persistent_column_feature()
if key == LANGUAGE_PERSISTENT
else language_events_column_feature()
)
elif ft["dtype"] == "video":
continue
elif ft["dtype"] == "image":
hf_features[key] = datasets.Image()
@@ -108,6 +122,41 @@ def create_empty_dataset_info(
)
def features_equal_for_merge(features_a: dict[str, dict], features_b: dict[str, dict]) -> bool:
"""Return whether two LeRobotDatasetMetadata ``features`` dicts are compatible for aggregation.
For video features, keys under ``info`` related to video encoding parameters are ignored during
comparison as they do not prevent aggregation.
"""
def _without_encoder_info_keys(feature: dict) -> dict:
filtered = dict(feature)
filtered_info = filtered.get("info")
if isinstance(filtered_info, dict):
filtered["info"] = {
info_key: info_value
for info_key, info_value in filtered_info.items()
if info_key not in VIDEO_ENCODER_INFO_KEYS
}
return filtered
if set(features_a) != set(features_b):
return False
for key in features_a:
fa_key = features_a[key]
fb_key = features_b[key]
if fa_key.get("dtype") != fb_key.get("dtype"):
return False
if fa_key.get("dtype") != "video":
if fa_key != fb_key:
return False
continue
if _without_encoder_info_keys(fa_key) != _without_encoder_info_keys(fb_key):
return False
return True
def check_delta_timestamps(
delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
) -> bool:
@@ -242,6 +291,8 @@ def validate_feature_dtype_and_shape(
return validate_feature_image_or_video(name, expected_shape, value)
elif expected_dtype == "string":
return validate_feature_string(name, value)
elif expected_dtype == "language":
return validate_feature_language(name, value)
else:
raise NotImplementedError(f"The feature dtype '{expected_dtype}' is not implemented yet.")
@@ -321,6 +372,30 @@ def validate_feature_string(name: str, value: str) -> str:
return ""
def validate_feature_language(name: str, value) -> str:
"""Validate a feature that is expected to hold language annotations.
Language columns (``language_persistent`` / ``language_events``) are
populated after recording by the annotation pipeline, not at record time.
Any value supplied here is dropped before the frame is written, so a
non-empty value almost certainly signals a mistake. We warn rather than
fail to keep recording resilient.
Args:
name (str): The name of the feature.
value: The value to validate.
Returns:
str: Always an empty string — language values are non-fatal.
"""
if value is not None:
logging.warning(
f"The feature '{name}' is a 'language' column populated by the annotation pipeline, "
f"not at record time. The provided value will be dropped."
)
return ""
def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features: dict) -> None:
"""Validate the episode buffer before it's written to disk.
src/lerobot/datasets/io_utils.py
+6 -11
View File
@@ -31,10 +31,10 @@ from torchvision import transforms
from lerobot.utils.io_utils import load_json, write_json
from lerobot.utils.utils import SuppressProgressBars, flatten_dict, unflatten_dict
from .language import LANGUAGE_COLUMNS
from .utils import (
DEFAULT_DATA_FILE_SIZE_IN_MB,
DEFAULT_EPISODES_PATH,
DEFAULT_SUBTASKS_PATH,
DEFAULT_TASKS_PATH,
EPISODES_DIR,
INFO_PATH,
@@ -186,14 +186,6 @@ def load_tasks(local_dir: Path) -> pandas.DataFrame:
return tasks
def load_subtasks(local_dir: Path) -> pandas.DataFrame | None:
"""Load subtasks from subtasks.parquet if it exists."""
subtasks_path = local_dir / DEFAULT_SUBTASKS_PATH
if subtasks_path.exists():
return pd.read_parquet(subtasks_path)
return None
def write_episodes(episodes: Dataset, local_dir: Path) -> None:
"""Write episode metadata to a parquet file in the LeRobot v3.0 format.
This function writes episode-level metadata to a single parquet file.
@@ -265,11 +257,13 @@ def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[to
dict: The batch with items converted to torch tensors.
"""
for key in items_dict:
if key in LANGUAGE_COLUMNS:
continue
first_item = items_dict[key][0]
if isinstance(first_item, PILImage.Image):
to_tensor = transforms.ToTensor()
items_dict[key] = [to_tensor(img) for img in items_dict[key]]
elif first_item is None:
elif first_item is None or isinstance(first_item, dict):
pass
else:
items_dict[key] = [x if isinstance(x, str) else torch.tensor(x) for x in items_dict[key]]
@@ -304,8 +298,9 @@ def item_to_torch(item: dict) -> dict:
Returns:
dict: Dictionary with all tensor-like items converted to torch.Tensor.
"""
skip_keys = {"task", *LANGUAGE_COLUMNS}
for key, val in item.items():
if isinstance(val, (np.ndarray | list)) and key not in ["task"]:
if isinstance(val, (np.ndarray | list)) and key not in skip_keys:
# Convert numpy arrays and lists to torch tensors
item[key] = torch.tensor(val)
return item
src/lerobot/datasets/language.py
+242
View File
@@ -0,0 +1,242 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations
from typing import Literal
import datasets
import pyarrow as pa
LANGUAGE_PERSISTENT = "language_persistent"
LANGUAGE_EVENTS = "language_events"
LANGUAGE_COLUMNS = (LANGUAGE_PERSISTENT, LANGUAGE_EVENTS)
PERSISTENT_ROW_FIELDS = ("role", "content", "style", "timestamp", "camera", "tool_calls")
EVENT_ROW_FIELDS = ("role", "content", "style", "camera", "tool_calls")
CORE_STYLES = {
"subtask",
"plan",
"memory",
"motion",
"interjection",
"vqa",
"trace",
"task_aug",
}
# Project-local styles can be registered at import time by appending to
# ``EXTENDED_STYLES`` before ``column_for_style`` is called. Anything added
# here is treated as a known style alongside ``CORE_STYLES`` for resolver
# validation. Empty by default — populate from a downstream module that
# also extends ``PERSISTENT_STYLES`` or ``EVENT_ONLY_STYLES`` to declare
# the new style's column.
EXTENDED_STYLES: set[str] = set()
STYLE_REGISTRY = CORE_STYLES | EXTENDED_STYLES
PERSISTENT_STYLES = {"subtask", "plan", "memory", "motion", "task_aug"}
EVENT_ONLY_STYLES = {"interjection", "vqa", "trace"}
# Styles whose ``content`` is grounded in a specific camera view. Rows of these
# styles MUST carry a non-null ``camera`` referencing an ``observation.images.*``
# feature key. Rows of every other style MUST have ``camera=None``. ``motion``
# is intentionally NOT in this set: motion primitives are described in
# robot-frame (joint / Cartesian) terms, not pixel space, so they are
# camera-agnostic. ``trace`` is the pixel-trajectory event style and IS
# view-dependent. The ``camera`` field nevertheless lives on
# ``PERSISTENT_ROW_FIELDS`` too so the schema, validator, and resolver
# behave symmetrically across the two columns; persistent rows simply
# always have ``camera=None`` in practice today.
VIEW_DEPENDENT_STYLES = {"vqa", "trace"}
LanguageColumn = Literal["language_persistent", "language_events"]
def _json_arrow_type() -> pa.DataType:
"""Return the Arrow JSON type, falling back to ``string`` on older pyarrow."""
return pa.json_() if hasattr(pa, "json_") else pa.string()
def _json_feature() -> object:
"""Return the HF ``datasets`` JSON feature, falling back to a string value."""
return datasets.Json() if hasattr(datasets, "Json") else datasets.Value("string")
def language_persistent_row_arrow_type() -> pa.StructType:
"""Return the Arrow struct type for a single persistent language row.
Persistent rows carry their own ``timestamp`` because they represent a state
that became active at a specific moment and remains active until superseded.
``timestamp`` is ``float32`` to match the timestamp dtype LeRobotDataset
uses for frame data.
"""
return pa.struct(
[
pa.field("role", pa.string(), nullable=False),
pa.field("content", pa.string(), nullable=True),
pa.field("style", pa.string(), nullable=True),
pa.field("timestamp", pa.float32(), nullable=False),
pa.field("camera", pa.string(), nullable=True),
pa.field("tool_calls", pa.list_(_json_arrow_type()), nullable=True),
]
)
def language_event_row_arrow_type() -> pa.StructType:
"""Return the Arrow struct type for a single event language row.
Event rows have no ``timestamp`` field: each event is stored on the dataset
row whose frame timestamp is the event's firing time.
"""
return pa.struct(
[
pa.field("role", pa.string(), nullable=False),
pa.field("content", pa.string(), nullable=True),
pa.field("style", pa.string(), nullable=True),
pa.field("camera", pa.string(), nullable=True),
pa.field("tool_calls", pa.list_(_json_arrow_type()), nullable=True),
]
)
def language_persistent_arrow_type() -> pa.ListType:
"""Return the Arrow list type for the ``language_persistent`` column."""
return pa.list_(language_persistent_row_arrow_type())
def language_events_arrow_type() -> pa.ListType:
"""Return the Arrow list type for the ``language_events`` column."""
return pa.list_(language_event_row_arrow_type())
def language_persistent_row_feature() -> dict[str, object]:
"""Return the HF ``datasets`` feature mapping for a persistent language row."""
return {
"role": datasets.Value("string"),
"content": datasets.Value("string"),
"style": datasets.Value("string"),
"timestamp": datasets.Value("float32"),
"camera": datasets.Value("string"),
"tool_calls": datasets.List(_json_feature()),
}
def language_event_row_feature() -> dict[str, object]:
"""Return the HF ``datasets`` feature mapping for an event language row."""
return {
"role": datasets.Value("string"),
"content": datasets.Value("string"),
"style": datasets.Value("string"),
"camera": datasets.Value("string"),
"tool_calls": datasets.List(_json_feature()),
}
def language_persistent_column_feature() -> datasets.List:
"""Return the HF ``datasets`` feature for the ``language_persistent`` column."""
return datasets.List(language_persistent_row_feature())
def language_events_column_feature() -> datasets.List:
"""Return the HF ``datasets`` feature for the ``language_events`` column."""
return datasets.List(language_event_row_feature())
def language_feature_info() -> dict[str, dict]:
"""Return the ``info["features"]`` entries for both language columns."""
return {
LANGUAGE_PERSISTENT: {"dtype": "language", "shape": (1,), "names": None},
LANGUAGE_EVENTS: {"dtype": "language", "shape": (1,), "names": None},
}
def is_language_column(key: str) -> bool:
"""Return ``True`` if ``key`` is one of the dataset's language column names."""
return key in LANGUAGE_COLUMNS
def is_view_dependent_style(style: str | None) -> bool:
"""Return ``True`` if rows of ``style`` must be tagged with a ``camera`` key."""
return style in VIEW_DEPENDENT_STYLES
def validate_camera_field(style: str | None, camera: str | None) -> None:
"""Enforce the ``camera`` invariant: required iff ``style`` is view-dependent.
Raises ``ValueError`` if a view-dependent style is missing ``camera`` or if
a non-view-dependent style carries one. Pipeline writers and the validator
should call this on every emitted row.
"""
if is_view_dependent_style(style):
if not camera:
raise ValueError(
f"Rows of view-dependent style {style!r} require a non-empty 'camera' "
f"field referencing an 'observation.images.*' feature key."
)
elif camera is not None:
raise ValueError(f"Rows of style {style!r} must have camera=None; got camera={camera!r}.")
# --- Tool registry --------------------------------------------------------
# Tools declared on a dataset live in ``meta/info.json["tools"]`` as a list
# of OpenAI-style function schemas. The runtime / training stack reads them
# through :class:`LeRobotDatasetMetadata.tools` (with these constants as
# fallback when the dataset doesn't declare any). Implementations live
# under :mod:`lerobot.tools` (one file per tool); see
# ``docs/source/tools.mdx`` for the authoring guide.
SAY_TOOL_SCHEMA: dict = {
"type": "function",
"function": {
"name": "say",
"description": "Speak a short utterance to the user via the TTS executor.",
"parameters": {
"type": "object",
"properties": {
"text": {
"type": "string",
"description": "The verbatim text to speak.",
}
},
"required": ["text"],
},
},
}
"""Canonical schema for the ``say`` tool emitted by the steerable
annotation pipeline (PR 2 Module 2). Single source of truth — PR 2's
writer, PR 3's runtime tool registry, and the dataset visualizer all
import this constant rather than duplicating the dict."""
DEFAULT_TOOLS: list[dict] = [SAY_TOOL_SCHEMA]
"""Fallback tools list. Returned by ``LeRobotDatasetMetadata.tools``
when ``meta/info.json["tools"]`` is unset, so unannotated datasets and
chat-template consumers (``apply_chat_template(messages, tools=...)``)
keep working out of the box."""
def column_for_style(style: str | None) -> LanguageColumn:
"""Map a language style to the column where rows of that style are stored.
Styles in :data:`PERSISTENT_STYLES` route to :data:`LANGUAGE_PERSISTENT`.
Styles in :data:`EVENT_ONLY_STYLES` and the implicit ``None`` style route
to :data:`LANGUAGE_EVENTS`.
"""
if style is None:
return LANGUAGE_EVENTS
if style in PERSISTENT_STYLES:
return LANGUAGE_PERSISTENT
if style in EVENT_ONLY_STYLES:
return LANGUAGE_EVENTS
raise ValueError(f"Unknown language style: {style!r}")
src/lerobot/datasets/language_render.py
+545
View File
@@ -0,0 +1,545 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations
import copy
import hashlib
import re
from collections.abc import Sequence
from typing import Any
from lerobot.configs.recipe import DEFAULT_BINDINGS, PLACEHOLDER_RE, TrainingRecipe
from lerobot.utils.utils import unwrap_scalar
from .language import LANGUAGE_PERSISTENT, column_for_style
LanguageRow = dict[str, Any]
RenderedMessages = dict[str, list[Any]]
_RESOLVER_RE = re.compile(r"^(?P<name>[A-Za-z_][A-Za-z0-9_]*)\((?P<args>.*)\)$")
def active_at(
t: float,
*,
persistent: Sequence[LanguageRow],
style: str | None = None,
role: str | None = None,
tool_name: str | None = None,
camera: str | None = None,
) -> LanguageRow | None:
"""Return the persistent row of ``style`` that is active at time ``t``.
A persistent row is "active" at ``t`` when its own ``timestamp`` is the
most recent one ``<= t`` for the given ``style``/``role``/``tool_name``/
``camera`` selector. Only valid for persistent styles.
"""
_validate_persistent_resolver("active_at", style)
matches = [
row
for row in _matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera)
if _timestamp(row) <= t
]
if not matches:
return None
latest_ts = max(_timestamp(row) for row in matches)
return _select_one(
[row for row in matches if _timestamp(row) == latest_ts],
style=style,
role=role,
tool_name=tool_name,
camera=camera,
)
EMITTED_AT_TOLERANCE_S = 0.1
"""Half-window for matching persistent rows to a frame timestamp in
``emitted_at``. Persistent timestamps come from parquet (float32) and ``t``
is also a float32 from parquet, so in the ideal hot path an exact match
would suffice — but any caller that derives ``t`` arithmetically (e.g.
``frame_idx / fps``) breaks bit-equality. A 0.1 s tolerance covers
common arithmetic drift without admitting frames that are visibly far
apart at typical control rates (30–100 Hz). This does mean two persistent
rows of the same selector emitted within 0.1 s of each other cannot be
told apart by ``emitted_at`` — acceptable because persistent annotations
(subtask / plan / memory transitions) change on a human-action timescale,
not at the camera frame rate."""
def emitted_at(
t: float,
*,
persistent: Sequence[LanguageRow],
events: Sequence[LanguageRow],
style: str | None = None,
role: str | None = None,
tool_name: str | None = None,
camera: str | None = None,
) -> LanguageRow | None:
"""Return the row of ``style`` emitted at exactly time ``t``.
For persistent styles, this matches persistent rows whose own ``timestamp``
is within ``EMITTED_AT_TOLERANCE_S`` of ``t`` (see that constant for why
we use a tolerance instead of bit-equality). For event styles, the
``events`` list is assumed to come from the dataset row at frame ``t``
(event rows carry no timestamp of their own), so all matching event rows
are considered emitted at ``t``. ``camera`` filters by the row's
``camera`` field — required to disambiguate when multiple view-dependent
rows share ``(t, role)`` across cameras.
"""
if column_for_style(style) == LANGUAGE_PERSISTENT:
matches = [
row
for row in _matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera)
if abs(_timestamp(row) - t) <= EMITTED_AT_TOLERANCE_S
]
else:
matches = _matching_rows(events, style=style, role=role, tool_name=tool_name, camera=camera)
return _select_one(matches, style=style, role=role, tool_name=tool_name, camera=camera)
def nth_prev(
t: float,
*,
persistent: Sequence[LanguageRow],
style: str | None = None,
offset: int = 1,
role: str | None = None,
tool_name: str | None = None,
camera: str | None = None,
) -> LanguageRow | None:
"""Return the persistent row that was active ``offset`` steps before ``t``.
Walks back through chronologically sorted persistent rows of ``style``
(filtered by optional ``role``/``tool_name``/``camera``) and returns the
one ``offset`` positions before the row active at ``t``. Only valid for
persistent styles.
"""
return _nth_relative("nth_prev", t, persistent, style, -offset, role, tool_name, camera)
def nth_next(
t: float,
*,
persistent: Sequence[LanguageRow],
style: str | None = None,
offset: int = 1,
role: str | None = None,
tool_name: str | None = None,
camera: str | None = None,
) -> LanguageRow | None:
"""Return the persistent row that becomes active ``offset`` steps after ``t``.
Walks forward through chronologically sorted persistent rows of ``style``
(filtered by optional ``role``/``tool_name``/``camera``) and returns the
one ``offset`` positions after the row active at ``t``. Only valid for
persistent styles.
"""
return _nth_relative("nth_next", t, persistent, style, offset, role, tool_name, camera)
def render_sample(
*,
recipe: TrainingRecipe,
persistent: Sequence[LanguageRow] | None,
events: Sequence[LanguageRow] | None,
t: float,
sample_idx: int,
task: str | None = None,
dataset_ctx: Any | None = None,
) -> RenderedMessages | None:
"""Render the chat-style messages for a single dataset sample.
Resolves the recipe's bindings against ``persistent`` and ``events`` rows
at frame timestamp ``t``, then expands the recipe's message templates.
Returns ``None`` if the resolved sample contains no target message.
"""
persistent_rows = _normalize_rows(persistent or [])
event_rows = _normalize_rows(events or [])
selected_recipe = _select_recipe(recipe, sample_idx)
bindings = _resolve_bindings(
selected_recipe,
persistent=persistent_rows,
events=event_rows,
t=t,
sample_idx=sample_idx,
task=task,
dataset_ctx=dataset_ctx,
)
return _render_message_recipe(selected_recipe, bindings)
def _select_recipe(recipe: TrainingRecipe, sample_idx: int) -> TrainingRecipe:
"""Pick a deterministic blend component for ``sample_idx`` (or return ``recipe``)."""
if recipe.blend is None:
return recipe
total_weight = sum(component.weight or 0.0 for component in recipe.blend.values())
if total_weight <= 0:
raise ValueError("Blend weights must sum to a positive value.")
digest = hashlib.blake2b(str(sample_idx).encode(), digest_size=8).digest()
draw = int.from_bytes(digest, "big") / 2**64 * total_weight
cumulative = 0.0
last_component: TrainingRecipe | None = None
for component in recipe.blend.values():
last_component = component
cumulative += component.weight or 0.0
if draw < cumulative:
return component
assert last_component is not None
return last_component
def _resolve_bindings(
recipe: TrainingRecipe,
*,
persistent: Sequence[LanguageRow],
events: Sequence[LanguageRow],
t: float,
sample_idx: int,
task: str | None,
dataset_ctx: Any | None,
) -> dict[str, LanguageRow | str | None]:
"""Resolve every binding in ``recipe`` (plus ``task``) at time ``t``."""
bindings: dict[str, LanguageRow | str | None] = {
"task": _resolve_task(task, dataset_ctx, persistent=persistent, sample_idx=sample_idx),
}
specs = {**DEFAULT_BINDINGS, **(recipe.bindings or {})}
for name, spec in specs.items():
bindings[name] = _resolve_spec(spec, persistent=persistent, events=events, t=t)
return bindings
def _resolve_task(
task: str | None,
dataset_ctx: Any | None,
*,
persistent: Sequence[LanguageRow] = (),
sample_idx: int = 0,
) -> str | None:
"""Return the task string for ``sample_idx``.
Resolution order:
1. Explicit ``task`` override (caller-supplied) wins.
2. If ``persistent`` contains rows of style ``task_aug`` (role=user),
deterministically pick one by ``sample_idx`` so each frame of an
episode rotates through the available rephrasings across an epoch.
This realizes Xiao 2022 / CAST-style task-prompt diversity without
changing ``meta/tasks.parquet`` and without forcing recipes to opt
in: ``${task}`` automatically picks a rephrasing when one exists,
and falls back to the canonical task otherwise. Recipes that want
the literal canonical task can override the binding.
3. Otherwise read the canonical task from ``dataset_ctx`` (which is
backed by ``meta/tasks.parquet``).
"""
if task is not None:
return task
aug_rows = [r for r in persistent if r.get("style") == "task_aug" and r.get("role") == "user"]
if aug_rows:
# Deterministic, blake2b-based pick keyed on sample_idx so the
# rotation is reproducible across runs (Python's built-in ``hash``
# is process-randomized).
digest = hashlib.blake2b(f"task_aug:{sample_idx}".encode(), digest_size=8).digest()
idx = int.from_bytes(digest, "big") % len(aug_rows)
chosen = aug_rows[idx].get("content")
if chosen:
return str(chosen)
if dataset_ctx is None:
return None
if isinstance(dataset_ctx, dict):
return dataset_ctx.get("task")
return getattr(dataset_ctx, "task", None)
def _resolve_spec(
spec: str,
*,
persistent: Sequence[LanguageRow],
events: Sequence[LanguageRow],
t: float,
) -> LanguageRow | None:
"""Parse a single binding's resolver expression and dispatch to its function."""
match = _RESOLVER_RE.match(spec.strip())
if match is None:
raise ValueError(f"Invalid resolver expression: {spec!r}")
name = match.group("name")
kwargs = _parse_resolver_args(match.group("args"))
kwargs.pop("t_arg", None)
if name == "emitted_at":
return emitted_at(t, persistent=persistent, events=events, **kwargs)
if name == "active_at":
return active_at(t, persistent=persistent, **kwargs)
if name == "nth_prev":
return nth_prev(t, persistent=persistent, **kwargs)
if name == "nth_next":
return nth_next(t, persistent=persistent, **kwargs)
raise ValueError(f"Unknown language resolver: {name!r}")
def _parse_resolver_args(args: str) -> dict[str, Any]:
"""Parse a comma-separated resolver argument list into a kwargs dict."""
kwargs: dict[str, Any] = {}
if not args.strip():
return kwargs
parts = [part.strip() for part in args.split(",") if part.strip()]
for part in parts:
if part == "t":
kwargs["t_arg"] = True
continue
if "=" not in part:
raise ValueError(f"Invalid resolver argument: {part!r}")
key, value = (item.strip() for item in part.split("=", 1))
if key == "offset":
kwargs[key] = int(value)
else:
kwargs[key] = value.strip("\"'")
return kwargs
def _render_message_recipe(
recipe: TrainingRecipe,
bindings: dict[str, LanguageRow | str | None],
) -> RenderedMessages | None:
"""Expand ``recipe.messages`` into rendered chat messages using ``bindings``."""
assert recipe.messages is not None
messages: list[dict[str, Any]] = []
streams: list[str | None] = []
target_indices: list[int] = []
for turn in recipe.messages:
if turn.if_present is not None and bindings.get(turn.if_present) is None:
continue
message = {"role": turn.role}
if turn.content is not None:
message["content"] = _render_content(turn.content, bindings)
if turn.tool_calls_from is not None:
row = bindings.get(turn.tool_calls_from)
tool_calls = row.get("tool_calls") if isinstance(row, dict) else None
if tool_calls:
message["tool_calls"] = copy.deepcopy(tool_calls)
message_idx = len(messages)
messages.append(message)
streams.append(turn.stream)
if turn.target:
target_indices.append(message_idx)
if not target_indices:
return None
rendered = {
"messages": messages,
"message_streams": streams,
"target_message_indices": target_indices,
}
_validate_rendered(rendered)
return rendered
def _render_content(
content: str | list[dict[str, Any]],
bindings: dict[str, LanguageRow | str | None],
) -> str | list[dict[str, Any]]:
"""Substitute bindings into a string or each string field of multimodal blocks."""
if isinstance(content, str):
return _substitute(content, bindings)
rendered_blocks = []
for block in content:
rendered_block = copy.deepcopy(block)
for key, value in rendered_block.items():
if isinstance(value, str):
rendered_block[key] = _substitute(value, bindings)
rendered_blocks.append(rendered_block)
return rendered_blocks
def _substitute(template: str, bindings: dict[str, LanguageRow | str | None]) -> str:
"""Replace ``${name}`` placeholders in ``template`` with their bound values."""
def replace(match: re.Match[str]) -> str:
"""Resolve a single ``${name}`` match to its bound string value."""
name = match.group(1)
if name not in bindings:
raise ValueError(f"Unknown template binding: {name!r}")
value = bindings[name]
if value is None:
return ""
if isinstance(value, dict):
content = value.get("content")
return "" if content is None else str(content)
return str(value)
return PLACEHOLDER_RE.sub(replace, template)
def _validate_rendered(rendered: RenderedMessages) -> None:
"""Sanity-check the rendered output for stream/target alignment."""
messages = rendered["messages"]
streams = rendered["message_streams"]
target_indices = rendered["target_message_indices"]
if len(streams) != len(messages):
raise ValueError("message_streams must be aligned with messages.")
if not target_indices:
raise ValueError("Rendered samples must contain at least one target message.")
for idx in target_indices:
if idx < 0 or idx >= len(messages):
raise ValueError(f"Target message index {idx} is out of bounds.")
# ``stream`` is enforced non-None at MessageTurn construction time
# (see ``MessageTurn.__post_init__``), so a missing stream here would
# mean the dataclass invariant was bypassed; no need to re-check.
def _nth_relative(
name: str,
t: float,
persistent: Sequence[LanguageRow],
style: str | None,
offset: int,
role: str | None,
tool_name: str | None,
camera: str | None,
) -> LanguageRow | None:
"""Shared body for ``nth_prev`` / ``nth_next`` with signed ``offset``."""
_validate_persistent_resolver(name, style)
if abs(offset) < 1:
raise ValueError(f"{name} offset must be non-zero.")
rows = sorted(
_matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera),
key=_row_sort_key,
)
if not rows:
return None
anchor_idx = None
for idx, row in enumerate(rows):
if _timestamp(row) <= t:
anchor_idx = idx
else:
break
target_idx = (offset - 1 if offset > 0 else None) if anchor_idx is None else anchor_idx + offset
if target_idx is None or target_idx < 0 or target_idx >= len(rows):
return None
return rows[target_idx]
def _validate_persistent_resolver(name: str, style: str | None) -> None:
"""Reject calls with missing or event-only ``style`` for persistent resolvers."""
if style is None:
raise ValueError(f"{name} requires a persistent style.")
if column_for_style(style) != LANGUAGE_PERSISTENT:
raise ValueError(f"{name} cannot be used with event-only style {style!r}.")
def _matching_rows(
rows: Sequence[LanguageRow],
*,
style: str | None,
role: str | None,
tool_name: str | None,
camera: str | None,
) -> list[LanguageRow]:
"""Return ``rows`` filtered by optional ``style``/``role``/``tool_name``/``camera`` selectors."""
return [
row
for row in rows
if (style is None or row.get("style") == style)
and (role is None or row.get("role") == role)
and (tool_name is None or _row_has_tool_name(row, tool_name))
and (camera is None or row.get("camera") == camera)
]
def _select_one(
rows: Sequence[LanguageRow],
*,
style: str | None,
role: str | None,
tool_name: str | None,
camera: str | None,
) -> LanguageRow | None:
"""Return the single matching row, or raise if the resolver is ambiguous.
Multiple matches always raise — even when the caller already passed
some selectors — because remaining ambiguity means the data has
several rows that look identical to the resolver and the caller
needs to pin down a specific one (e.g. add ``camera=...`` for VQA
rows shared across cameras).
"""
if not rows:
return None
if len(rows) > 1:
raise ValueError(
f"Ambiguous resolver for style={style!r} role={role!r} "
f"tool_name={tool_name!r} camera={camera!r}: {len(rows)} matching rows. "
f"Add a selector that distinguishes them."
)
return rows[0]
def _row_sort_key(row: LanguageRow) -> tuple[float, str, str]:
"""Stable sort key for both persistent and event rows.
Event rows lack ``timestamp`` (it is implicit in the frame), so default
to ``0.0`` — within a single frame all event rows share the same sort
bucket and are tiebroken by ``(style, role)``.
"""
timestamp = row.get("timestamp")
ts = float(unwrap_scalar(timestamp)) if timestamp is not None else 0.0
return (ts, row.get("style") or "", row.get("role") or "")
def _timestamp(row: LanguageRow) -> float:
"""Extract a row's ``timestamp`` as a Python float (unwrapping numpy scalars)."""
return float(unwrap_scalar(row["timestamp"]))
def _row_has_tool_name(row: LanguageRow, tool_name: str) -> bool:
"""Return ``True`` if any of the row's tool calls invokes ``tool_name``."""
for tool_call in row.get("tool_calls") or []:
if isinstance(tool_call, str):
continue
function = tool_call.get("function") if isinstance(tool_call, dict) else None
if isinstance(function, dict) and function.get("name") == tool_name:
return True
return False
def _normalize_rows(rows: Sequence[Any]) -> list[LanguageRow]:
"""Convert pyarrow scalars / mappings into a fresh list of plain dict rows."""
normalized = []
for row in rows:
if row is None:
continue
if hasattr(row, "as_py"):
row = row.as_py()
if not isinstance(row, dict):
raise TypeError(f"Language rows must be dictionaries, got {type(row).__name__}.")
normalized.append(dict(row))
return normalized
src/lerobot/datasets/lerobot_dataset.py
+38 -41
View File
@@ -24,6 +24,7 @@ import torch.utils
from huggingface_hub import HfApi, snapshot_download
from huggingface_hub.errors import RevisionNotFoundError
from lerobot.configs import VideoEncoderConfig
from lerobot.utils.constants import HF_LEROBOT_HUB_CACHE
from .dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
@@ -36,8 +37,7 @@ from .utils import (
)
from .video_utils import (
StreamingVideoEncoder,
get_safe_default_codec,
resolve_vcodec,
get_safe_default_video_backend,
)
logger = logging.getLogger(__name__)
@@ -59,10 +59,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
video_backend: str | None = None,
return_uint8: bool = False,
batch_encoding_size: int = 1,
vcodec: str = "libsvtav1",
camera_encoder: VideoEncoderConfig | None = None,
encoder_threads: int | None = None,
streaming_encoding: bool = False,
encoder_queue_maxsize: int = 30,
encoder_threads: int | None = None,
):
"""
2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -183,16 +183,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
You can also use the 'pyav' decoder used by Torchvision, which used to be the default option, or 'video_reader' which is another decoder of Torchvision.
batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
vcodec (str, optional): Video codec for encoding videos during recording. Options: 'h264', 'hevc',
'libsvtav1', 'auto', or hardware-specific codecs like 'h264_videotoolbox', 'h264_nvenc'.
Defaults to 'libsvtav1'. Use 'auto' to auto-detect the best available hardware encoder.
camera_encoder (VideoEncoderConfig | None, optional): Video encoder settings for cameras
(codec, quality, etc.). When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults`
is used by the writer.
encoder_threads (int | None, optional): Number of encoder threads (global). ``None`` lets the
codec decide.
streaming_encoding (bool, optional): If True, encode video frames in real-time during capture
instead of writing PNG images first. This makes save_episode() near-instant. Defaults to False.
encoder_queue_maxsize (int, optional): Maximum number of frames to buffer per camera when using
streaming encoding. Defaults to 30 (~1s at 30fps).
encoder_threads (int | None, optional): Number of threads per encoder instance. None lets the
codec auto-detect (default). Lower values reduce CPU usage per encoder. Maps to 'lp' (via svtav1-params) for
libsvtav1 and 'threads' for h264/hevc.
Note:
Write-mode parameters (``streaming_encoding``, ``batch_encoding_size``) passed to
@@ -207,10 +206,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
self.delta_timestamps = delta_timestamps
self.tolerance_s = tolerance_s
self.revision = revision if revision else CODEBASE_VERSION
self._video_backend = video_backend if video_backend else get_safe_default_codec()
self._video_backend = video_backend if video_backend else get_safe_default_video_backend()
self._return_uint8 = return_uint8
self._batch_encoding_size = batch_encoding_size
self._vcodec = resolve_vcodec(vcodec)
self._encoder_threads = encoder_threads
if self._requested_root is not None:
@@ -273,12 +271,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
streaming_enc = None
if streaming_encoding and len(self.meta.video_keys) > 0:
streaming_enc = self._build_streaming_encoder(
self.meta.fps, self._vcodec, encoder_queue_maxsize, encoder_threads
self.meta.fps,
camera_encoder,
encoder_queue_maxsize,
encoder_threads,
)
self.writer = DatasetWriter(
meta=self.meta,
root=self.root,
vcodec=self._vcodec,
camera_encoder=camera_encoder,
encoder_threads=encoder_threads,
batch_encoding_size=batch_encoding_size,
streaming_encoder=streaming_enc,
@@ -320,17 +321,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
@staticmethod
def _build_streaming_encoder(
fps: int,
vcodec: str,
camera_encoder: VideoEncoderConfig | None,
encoder_queue_maxsize: int,
encoder_threads: int | None,
) -> StreamingVideoEncoder:
return StreamingVideoEncoder(
fps=fps,
vcodec=vcodec,
pix_fmt="yuv420p",
g=2,
crf=30,
preset=None,
camera_encoder=camera_encoder,
queue_maxsize=encoder_queue_maxsize,
encoder_threads=encoder_threads,
)
@@ -647,7 +644,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
image_writer_threads: int = 0,
video_backend: str | None = None,
batch_encoding_size: int = 1,
vcodec: str = "libsvtav1",
camera_encoder: VideoEncoderConfig | None = None,
metadata_buffer_size: int = 10,
streaming_encoding: bool = False,
encoder_queue_maxsize: int = 30,
@@ -678,20 +675,20 @@ class LeRobotDataset(torch.utils.data.Dataset):
video_backend: Video decoding backend (used when reading back).
batch_encoding_size: Number of episodes to accumulate before
batch-encoding videos. ``1`` means encode immediately.
vcodec: Video codec for encoding. Options include ``'libsvtav1'``,
``'h264'``, ``'hevc'``, ``'auto'``.
camera_encoder: Video encoder settings for cameras (codec, quality, etc.).
When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults` is used.
encoder_threads: Number of encoder threads (global). ``None``
lets the codec decide.
metadata_buffer_size: Number of episode metadata records to buffer
before flushing to parquet.
streaming_encoding: If ``True``, encode video frames in real-time
during capture instead of writing images first.
encoder_queue_maxsize: Max buffered frames per camera when using
streaming encoding.
encoder_threads: Threads per encoder instance. ``None`` for auto.
Returns:
A new :class:`LeRobotDataset` in write mode.
"""
vcodec = resolve_vcodec(vcodec)
obj = cls.__new__(cls)
obj.meta = LeRobotDatasetMetadata.create(
repo_id=repo_id,
@@ -712,23 +709,23 @@ class LeRobotDataset(torch.utils.data.Dataset):
obj.image_transforms = None
obj.delta_timestamps = None
obj.episodes = None
obj._video_backend = video_backend if video_backend is not None else get_safe_default_codec()
obj._video_backend = video_backend if video_backend is not None else get_safe_default_video_backend()
obj._return_uint8 = False
obj._batch_encoding_size = batch_encoding_size
obj._vcodec = vcodec
obj._encoder_threads = encoder_threads
# Reader is lazily created on first access (write-only mode)
obj.reader = None
# Create writer
streaming_enc = None
if streaming_encoding and len(obj.meta.video_keys) > 0:
streaming_enc = cls._build_streaming_encoder(fps, vcodec, encoder_queue_maxsize, encoder_threads)
streaming_enc = cls._build_streaming_encoder(
fps, camera_encoder, encoder_queue_maxsize, encoder_threads
)
obj.writer = DatasetWriter(
meta=obj.meta,
root=obj.root,
vcodec=vcodec,
camera_encoder=camera_encoder,
encoder_threads=encoder_threads,
batch_encoding_size=batch_encoding_size,
streaming_encoder=streaming_enc,
@@ -751,12 +748,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
force_cache_sync: bool = False,
video_backend: str | None = None,
batch_encoding_size: int = 1,
vcodec: str = "libsvtav1",
camera_encoder: VideoEncoderConfig | None = None,
encoder_threads: int | None = None,
image_writer_processes: int = 0,
image_writer_threads: int = 0,
streaming_encoding: bool = False,
encoder_queue_maxsize: int = 30,
encoder_threads: int | None = None,
) -> "LeRobotDataset":
"""Resume recording on an existing dataset.
@@ -779,13 +776,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
video_backend: Video decoding backend for reading back data.
batch_encoding_size: Number of episodes to accumulate before
batch-encoding videos.
vcodec: Video codec for encoding.
camera_encoder: Video encoder settings for cameras (codec, quality, etc.).
When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults` is used.
encoder_threads: Number of encoder threads (global). ``None``
lets the codec decide.
image_writer_processes: Subprocesses for async image writing.
image_writer_threads: Threads for async image writing.
streaming_encoding: If ``True``, encode video in real-time during
capture.
encoder_queue_maxsize: Max buffered frames per camera for streaming.
encoder_threads: Threads per encoder instance. ``None`` for auto.
Returns:
A :class:`LeRobotDataset` in write mode, ready to append episodes.
@@ -796,7 +795,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
"Writing into the revision-safe Hub snapshot cache (used when root=None) would corrupt "
"the shared cache. Please provide a local directory path."
)
vcodec = resolve_vcodec(vcodec)
obj = cls.__new__(cls)
obj.repo_id = repo_id
obj._requested_root = Path(root)
@@ -805,11 +803,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
obj.image_transforms = None
obj.delta_timestamps = None
obj.episodes = None
obj._video_backend = video_backend if video_backend else get_safe_default_codec()
obj._video_backend = video_backend if video_backend else get_safe_default_video_backend()
obj._return_uint8 = False
obj._batch_encoding_size = batch_encoding_size
obj._vcodec = vcodec
obj._encoder_threads = encoder_threads
if obj._requested_root is not None:
obj._requested_root.mkdir(exist_ok=True, parents=True)
@@ -818,21 +814,22 @@ class LeRobotDataset(torch.utils.data.Dataset):
obj.meta = LeRobotDatasetMetadata(
obj.repo_id, obj._requested_root, obj.revision, force_cache_sync=force_cache_sync
)
obj._encoder_threads = encoder_threads
obj.root = obj.meta.root
# Reader is lazily created on first access (write-only mode)
obj.reader = None
# Create writer for appending
streaming_enc = None
if streaming_encoding and len(obj.meta.video_keys) > 0:
streaming_enc = cls._build_streaming_encoder(
obj.meta.fps, vcodec, encoder_queue_maxsize, encoder_threads
obj.meta.fps, camera_encoder, encoder_queue_maxsize, encoder_threads
)
obj.writer = DatasetWriter(
meta=obj.meta,
root=obj.root,
vcodec=vcodec,
camera_encoder=camera_encoder,
encoder_threads=encoder_threads,
batch_encoding_size=batch_encoding_size,
streaming_encoder=streaming_enc,
src/lerobot/datasets/pyav_utils.py
+174
View File
@@ -0,0 +1,174 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyAV-based compatibility checks for :class:`VideoEncoderConfig`.
Centralises all :mod:`av` introspection of the bundled FFmpeg build.
Checks degrade to a no-op when the target codec isn't available locally.
"""
import functools
import logging
from typing import Any
import av
logger = logging.getLogger(__name__)
FFMPEG_NUMERIC_OPTION_TYPES = ("INT", "INT64", "UINT64", "FLOAT", "DOUBLE")
FFMPEG_INTEGER_OPTION_TYPES = ("INT", "INT64", "UINT64")
@functools.cache
def get_codec(vcodec: str) -> av.codec.Codec | None:
"""PyAV write-mode ``Codec`` for *vcodec*, or ``None`` if unavailable."""
try:
return av.codec.Codec(vcodec, "w")
except Exception:
return None
@functools.cache
def _get_codec_options_by_name(vcodec: str) -> dict[str, av.option.Option]:
"""Private-option name → PyAV ``Option`` for *vcodec* (empty if unavailable)."""
codec = get_codec(vcodec)
if codec is None:
return {}
return {opt.name: opt for opt in codec.descriptor.options}
@functools.cache
def _get_codec_video_formats(vcodec: str) -> tuple[str, ...]:
"""Pixel formats accepted by *vcodec* in PyAV's preferred order (empty if unknown)."""
codec = get_codec(vcodec)
if codec is None:
return ()
return tuple(fmt.name for fmt in (codec.video_formats or []))
def detect_available_encoders_pyav(encoders: list[str] | str) -> list[str]:
"""Return the subset of *encoders* available as video encoders in the local FFmpeg build.
Each name is probed directly via :func:`get_codec`; input order is preserved.
"""
if isinstance(encoders, str):
encoders = [encoders]
available: list[str] = []
for name in encoders:
codec = get_codec(name)
if codec is not None and codec.type == "video":
available.append(name)
else:
logger.debug("encoder '%s' not available as video encoder", name)
return available
def _check_option_value(vcodec: str, label: str, value: Any, opt: av.option.Option) -> None:
"""Range-check numeric *value* and choice-check string *value* against *opt*."""
type_name = opt.type.name
if type_name in FFMPEG_NUMERIC_OPTION_TYPES:
if isinstance(value, bool):
raise ValueError(
f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option."
)
elif isinstance(value, str):
try:
num_val = float(value)
except ValueError as e:
raise ValueError(
f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option."
) from e
elif isinstance(value, (float, int)):
num_val = value
else:
raise ValueError(
f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option."
)
# Check integer type compatibility
if type_name in FFMPEG_INTEGER_OPTION_TYPES and not num_val.is_integer():
raise ValueError(
f"{label}={num_val!r} must be an integer for codec {vcodec!r} "
f"(FFmpeg option {opt.name!r} is {type_name}); float values are not allowed."
)
# Check numeric range compatibility
lo, hi = float(opt.min), float(opt.max)
if lo < hi and not (lo <= num_val <= hi):
raise ValueError(
f"{label}={num_val} is out of range for codec {vcodec!r}; must be in [{lo}, {hi}]"
)
elif type_name == "STRING":
if isinstance(value, bool):
raise ValueError(f"{label}={value!r} is not a valid string value for codec {vcodec!r}.")
if isinstance(value, str):
str_val = value
elif isinstance(value, (int, float)):
str_val = str(value)
else:
raise ValueError(f"{label}={value!r} has unsupported type for STRING option on codec {vcodec!r}")
# Check string choice compatibility
choices = [c.name for c in (opt.choices or [])]
if choices and str_val not in choices:
raise ValueError(
f"{label}={str_val!r} is not a supported choice for codec "
f"{vcodec!r}; valid choices: {choices}"
)
else:
return
def _check_pixel_format(vcodec: str, pix_fmt: str) -> None:
formats = _get_codec_video_formats(vcodec)
if formats and pix_fmt not in formats:
raise ValueError(
f"pix_fmt={pix_fmt!r} is not supported by codec {vcodec!r}; "
f"supported pixel formats: {list(formats)}"
)
def _check_codec_options(vcodec: str, codec_options: dict[str, Any]) -> None:
"""Validate merged encoder options (typed) against the codec's published AVOptions."""
supported_options = _get_codec_options_by_name(vcodec)
for key, value in codec_options.items():
# GOP size is not a codec-specific option, it has to be validated separately.
if key == "g":
if isinstance(value, bool) or not isinstance(value, int) or value < 1:
raise ValueError(f"g={value!r} must be a positive integer for codec {vcodec!r}")
continue
if key not in supported_options:
continue
_check_option_value(vcodec, key, value, supported_options[key])
def check_video_encoder_parameters_pyav(vcodec: str, pix_fmt: str, codec_options: dict[str, Any]) -> None:
"""Verify *config* is compatible with the bundled FFmpeg build.
Checks pixel format, abstract tuning-field compatibility, and each merged
encoder option from :meth:`~lerobot.configs.video.VideoEncoderConfig.get_codec_options`
against PyAV (including numeric ``extra_options`` present in that dict).
No-op when ``config.vcodec`` isn't in the local FFmpeg build.
Raises:
ValueError: on the first incompatibility encountered.
"""
options = _get_codec_options_by_name(vcodec)
if not options:
raise ValueError(f"Codec {vcodec!r} is not available in the bundled FFmpeg build")
_check_pixel_format(vcodec, pix_fmt)
_check_codec_options(vcodec, codec_options)
src/lerobot/datasets/utils.py
+7 -1
View File
@@ -88,7 +88,6 @@ VIDEO_DIR = "videos"
CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
DEFAULT_TASKS_PATH = "meta/tasks.parquet"
DEFAULT_SUBTASKS_PATH = "meta/subtasks.parquet"
DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4"
@@ -130,6 +129,9 @@ class DatasetInfo:
# Optional metadata
robot_type: str | None = None
splits: dict[str, str] = field(default_factory=dict)
# OpenAI-style tool schemas declared by the dataset. ``None`` means the
# dataset doesn't declare any — readers fall back to ``DEFAULT_TOOLS``.
tools: list[dict] | None = None
def __post_init__(self) -> None:
# Coerce feature shapes from list to tuple — JSON deserialisation
@@ -151,11 +153,15 @@ class DatasetInfo:
"""Return a JSON-serialisable dict.
Converts tuple shapes back to lists so ``json.dump`` can handle them.
Drops ``tools`` when unset so existing datasets keep a clean
``info.json``.
"""
d = dataclasses.asdict(self)
for ft in d["features"].values():
if isinstance(ft.get("shape"), tuple):
ft["shape"] = list(ft["shape"])
if d.get("tools") is None:
d.pop("tools", None)
return d
@classmethod
src/lerobot/datasets/video_utils.py
+250 -167
View File
@@ -17,12 +17,14 @@ import contextlib
import glob
import importlib
import logging
import os
import queue
import shutil
import tempfile
import threading
import warnings
from dataclasses import dataclass, field
from collections import OrderedDict
from dataclasses import asdict, dataclass, field
from fractions import Fraction
from pathlib import Path
from threading import Lock
@@ -36,86 +38,14 @@ import torch
from datasets.features.features import register_feature
from PIL import Image
from lerobot.utils.import_utils import get_safe_default_codec
from lerobot.configs import (
VideoEncoderConfig,
camera_encoder_defaults,
)
from lerobot.utils.import_utils import get_safe_default_video_backend
logger = logging.getLogger(__name__)
# List of hardware encoders to probe for auto-selection. Availability depends on the platform and FFmpeg build.
# Determines the order of preference for auto-selection when vcodec="auto" is used.
HW_ENCODERS = [
"h264_videotoolbox", # macOS
"hevc_videotoolbox", # macOS
"h264_nvenc", # NVIDIA GPU
"hevc_nvenc", # NVIDIA GPU
"h264_vaapi", # Linux Intel/AMD
"h264_qsv", # Intel Quick Sync
]
VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1", "auto"} | set(HW_ENCODERS)
def _get_codec_options(
vcodec: str,
g: int | None = 2,
crf: int | None = 30,
preset: int | None = None,
) -> dict:
"""Build codec-specific options dict for video encoding."""
options = {}
# GOP size (keyframe interval) - supported by VideoToolbox and software encoders
if g is not None and (vcodec in ("h264_videotoolbox", "hevc_videotoolbox") or vcodec not in HW_ENCODERS):
options["g"] = str(g)
# Quality control (codec-specific parameter names)
if crf is not None:
if vcodec in ("h264", "hevc", "libsvtav1"):
options["crf"] = str(crf)
elif vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
quality = max(1, min(100, int(100 - crf * 2)))
options["q:v"] = str(quality)
elif vcodec in ("h264_nvenc", "hevc_nvenc"):
options["rc"] = "constqp"
options["qp"] = str(crf)
elif vcodec in ("h264_vaapi",):
options["qp"] = str(crf)
elif vcodec in ("h264_qsv",):
options["global_quality"] = str(crf)
# Preset (only for libsvtav1)
if vcodec == "libsvtav1":
options["preset"] = str(preset) if preset is not None else "12"
return options
def detect_available_hw_encoders() -> list[str]:
"""Probe PyAV/FFmpeg for available hardware video encoders."""
available = []
for codec_name in HW_ENCODERS:
try:
av.codec.Codec(codec_name, "w")
available.append(codec_name)
except Exception: # nosec B110
logger.debug("HW encoder '%s' not available", codec_name) # nosec B110
return available
def resolve_vcodec(vcodec: str) -> str:
"""Validate vcodec and resolve 'auto' to best available HW encoder, fallback to libsvtav1."""
if vcodec not in VALID_VIDEO_CODECS:
raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
if vcodec != "auto":
logger.info(f"Using video codec: {vcodec}")
return vcodec
available = detect_available_hw_encoders()
for encoder in HW_ENCODERS:
if encoder in available:
logger.info(f"Auto-selected video codec: {encoder}")
return encoder
logger.info("No hardware encoder available, falling back to software encoder 'libsvtav1'")
return "libsvtav1"
def decode_video_frames(
video_path: Path | str,
@@ -143,7 +73,7 @@ def decode_video_frames(
Currently supports torchcodec on cpu and pyav.
"""
if backend is None:
backend = get_safe_default_codec()
backend = get_safe_default_video_backend()
if backend == "torchcodec":
return decode_video_frames_torchcodec(video_path, timestamps, tolerance_s, return_uint8=return_uint8)
elif backend == "pyav":
@@ -263,15 +193,70 @@ def decode_video_frames_pyav(
return closest_frames
class VideoDecoderCache:
"""Thread-safe cache for video decoders to avoid expensive re-initialization."""
DEFAULT_DECODER_CACHE_SIZE = 100
"""Default LRU capacity for :class:`VideoDecoderCache`.
def __init__(self):
self._cache: dict[str, tuple[Any, Any]] = {}
Sized to comfortably hold a small rolling window of episodes worth of decoders
(typical recipes: 2-4 cameras per episode × tens of episodes in flight) while
bounding host RAM. Each cached entry retains a torchcodec ``VideoDecoder`` plus
an open ``fsspec`` file handle — on the order of a few MB per entry. Override
via the ``LEROBOT_VIDEO_DECODER_CACHE_SIZE`` env var or by passing ``max_size``
to the constructor (``None`` restores the legacy unbounded behaviour).
"""
def _default_max_cache_size() -> int | None:
raw = os.environ.get("LEROBOT_VIDEO_DECODER_CACHE_SIZE")
if raw is None:
return DEFAULT_DECODER_CACHE_SIZE
raw = raw.strip().lower()
if raw in ("", "none", "unbounded", "-1"):
return None
try:
value = int(raw)
except ValueError as e:
raise ValueError(
f"LEROBOT_VIDEO_DECODER_CACHE_SIZE must be an integer, 'none', or '-1'; got {raw!r}"
) from e
if value <= 0:
raise ValueError(f"LEROBOT_VIDEO_DECODER_CACHE_SIZE must be positive; got {value}")
return value
class VideoDecoderCache:
"""Thread-safe LRU cache for torchcodec ``VideoDecoder`` instances.
Cached entries hold a ``VideoDecoder`` plus the open ``fsspec`` file handle
backing it. When the cache is full and a new path is requested, the
least-recently-used entry is evicted and its file handle is closed. This
bounds host-RAM growth when iterating over datasets with many distinct
video files (otherwise each ``DataLoader`` worker pins every decoder it has
ever opened until the process exits).
Args:
max_size: Maximum number of decoders to retain. ``None`` disables
eviction and restores legacy unbounded behaviour. Defaults to the
value of ``LEROBOT_VIDEO_DECODER_CACHE_SIZE`` if set, otherwise
:data:`DEFAULT_DECODER_CACHE_SIZE`.
"""
_SENTINEL: ClassVar[object] = object()
def __init__(self, max_size: int | None | object = _SENTINEL):
if max_size is VideoDecoderCache._SENTINEL:
max_size = _default_max_cache_size()
if max_size is not None and max_size <= 0:
raise ValueError(f"max_size must be positive or None; got {max_size}")
self.max_size: int | None = max_size # type: ignore[assignment]
self._cache: OrderedDict[str, tuple[Any, Any]] = OrderedDict()
self._lock = Lock()
def __contains__(self, video_path: object) -> bool:
with self._lock:
return str(video_path) in self._cache
def get_decoder(self, video_path: str):
"""Get a cached decoder or create a new one."""
"""Get a cached decoder or create a new one, evicting LRU if at capacity."""
if importlib.util.find_spec("torchcodec"):
from torchcodec.decoders import VideoDecoder
else:
@@ -283,22 +268,36 @@ class VideoDecoderCache:
video_path = str(video_path)
with self._lock:
if video_path not in self._cache:
file_handle = fsspec.open(video_path).__enter__()
try:
decoder = VideoDecoder(file_handle, seek_mode="approximate")
except Exception:
file_handle.close()
raise
self._cache[video_path] = (decoder, file_handle)
entry = self._cache.get(video_path)
if entry is not None:
self._cache.move_to_end(video_path)
return entry[0]
return self._cache[video_path][0]
file_handle = fsspec.open(video_path).__enter__()
try:
decoder = VideoDecoder(file_handle, seek_mode="approximate")
except Exception:
file_handle.close()
raise
self._cache[video_path] = (decoder, file_handle)
# Evict LRU entries until we are back under the cap. We close
# evicted file handles immediately; the associated ``VideoDecoder``
# is released to the GC when its last reference goes away.
if self.max_size is not None:
while len(self._cache) > self.max_size:
_evicted_path, (_evicted_decoder, evicted_handle) = self._cache.popitem(last=False)
with contextlib.suppress(Exception):
evicted_handle.close()
return decoder
def clear(self):
"""Clear the cache and close file handles."""
"""Clear the cache and close all file handles."""
with self._lock:
for _, file_handle in self._cache.values():
file_handle.close()
with contextlib.suppress(Exception):
file_handle.close()
self._cache.clear()
def size(self) -> int:
@@ -407,18 +406,17 @@ def encode_video_frames(
imgs_dir: Path | str,
video_path: Path | str,
fps: int,
vcodec: str = "libsvtav1",
pix_fmt: str = "yuv420p",
g: int | None = 2,
crf: int | None = 30,
fast_decode: int = 0,
camera_encoder: VideoEncoderConfig | None = None,
encoder_threads: int | None = None,
*,
log_level: int | None = av.logging.WARNING,
overwrite: bool = False,
preset: int | None = None,
encoder_threads: int | None = None,
) -> None:
"""More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
vcodec = resolve_vcodec(vcodec)
if camera_encoder is None:
camera_encoder = camera_encoder_defaults()
vcodec = camera_encoder.vcodec
pix_fmt = camera_encoder.pix_fmt
video_path = Path(video_path)
imgs_dir = Path(imgs_dir)
@@ -429,42 +427,18 @@ def encode_video_frames(
video_path.parent.mkdir(parents=True, exist_ok=True)
# Encoders/pixel formats incompatibility check
if (vcodec == "libsvtav1" or vcodec == "hevc") and pix_fmt == "yuv444p":
logger.warning(
f"Incompatible pixel format 'yuv444p' for codec {vcodec}, auto-selecting format 'yuv420p'"
)
pix_fmt = "yuv420p"
# Get input frames
template = "frame-" + ("[0-9]" * 6) + ".png"
input_list = sorted(
glob.glob(str(imgs_dir / template)), key=lambda x: int(x.split("-")[-1].split(".")[0])
)
# Define video output frame size (assuming all input frames are the same size)
if len(input_list) == 0:
raise FileNotFoundError(f"No images found in {imgs_dir}.")
with Image.open(input_list[0]) as dummy_image:
width, height = dummy_image.size
# Define video codec options
video_options = _get_codec_options(vcodec, g, crf, preset)
if fast_decode:
key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
video_options[key] = value
if encoder_threads is not None:
if vcodec == "libsvtav1":
lp_param = f"lp={encoder_threads}"
if "svtav1-params" in video_options:
video_options["svtav1-params"] += f":{lp_param}"
else:
video_options["svtav1-params"] = lp_param
else:
video_options["threads"] = str(encoder_threads)
video_options = camera_encoder.get_codec_options(encoder_threads, as_strings=True)
# Set logging level
if log_level is not None:
@@ -500,8 +474,97 @@ def encode_video_frames(
raise OSError(f"Video encoding did not work. File not found: {video_path}.")
def reencode_video(
input_video_path: Path | str,
output_video_path: Path | str,
camera_encoder: VideoEncoderConfig | None = None,
encoder_threads: int | None = None,
log_level: int | None = av.logging.WARNING,
overwrite: bool = False,
) -> None:
"""Re-encode a video file using the given encoder configuration.
Args:
input_video_path: Existing video file to read.
output_video_path: Path for the re-encoded file.
camera_encoder: Encoder configuration. Defaults to :func:`camera_encoder_defaults`.
encoder_threads: Optional thread count forwarded to :meth:`VideoEncoderConfig.get_codec_options`.
log_level: libav log level while encoding, or ``None`` to leave logging unchanged. Defaults to WARNING.
overwrite: When ``False`` and ``output_video_path`` already exists, skip and log a warning.
"""
camera_encoder = camera_encoder or camera_encoder_defaults()
output_video_path = Path(output_video_path)
if output_video_path.exists() and not overwrite:
logger.warning(f"Video file already exists: {output_video_path}. Skipping re-encode.")
return
output_video_path.parent.mkdir(parents=True, exist_ok=True)
video_options = camera_encoder.get_codec_options(encoder_threads, as_strings=True)
vcodec = camera_encoder.vcodec
pix_fmt = camera_encoder.pix_fmt
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_named_file:
tmp_output_video_path = tmp_named_file.name
if log_level is not None:
logging.getLogger("libav").setLevel(log_level)
try:
with av.open(input_video_path, mode="r") as src:
try:
in_stream = src.streams.video[0]
except IndexError as e:
raise ValueError(f"No video stream in {input_video_path}") from e
fps = (
in_stream.base_rate
) # We allow fractional fps though LeRobotDataset only supports integer fps
width = int(in_stream.width)
height = int(in_stream.height)
with av.open(
tmp_output_video_path,
mode="w",
options={
"movflags": "faststart"
}, # faststart is to move the metadata to the beginning of the file to speed up loading
) as dst:
out_stream = dst.add_stream(vcodec, fps, options=video_options)
out_stream.pix_fmt = pix_fmt
out_stream.width = width
out_stream.height = height
for frame in src.decode(in_stream):
frame = frame.reformat(width=width, height=height, format=pix_fmt)
packet = out_stream.encode(frame)
if packet:
dst.mux(packet)
packet = out_stream.encode()
if packet:
dst.mux(packet)
shutil.move(tmp_output_video_path, output_video_path)
except Exception:
Path(tmp_output_video_path).unlink(missing_ok=True)
raise
finally:
if log_level is not None:
av.logging.restore_default_callback()
if not output_video_path.exists():
raise OSError(f"Video re-encoding did not work. File not found: {output_video_path}.")
def concatenate_video_files(
input_video_paths: list[Path | str], output_video_path: Path, overwrite: bool = True
input_video_paths: list[Path | str],
output_video_path: Path,
overwrite: bool = True,
compatibility_check: bool = False,
):
"""
Concatenate multiple video files into a single video file using pyav.
@@ -514,6 +577,7 @@ def concatenate_video_files(
input_video_paths: Ordered list of input video file paths to concatenate.
output_video_path: Path to the output video file.
overwrite: Whether to overwrite the output video file if it already exists. Default is True.
compatibility_check: Whether to check if the input videos are compatible. Default is False.
Note:
- Creates a temporary directory for intermediate files that is cleaned up after use.
@@ -532,6 +596,22 @@ def concatenate_video_files(
if len(input_video_paths) == 0:
raise FileNotFoundError("No input video paths provided.")
# This check may be skipped at recording time as videos are encoded with the same encoder config.
if compatibility_check:
reference_video_info = get_video_info(input_video_paths[0])
for input_path in input_video_paths[1:]:
video_info = get_video_info(input_path)
if (
video_info["video.height"] != reference_video_info["video.height"]
or video_info["video.width"] != reference_video_info["video.width"]
or video_info["video.fps"] != reference_video_info["video.fps"]
or video_info["video.codec"] != reference_video_info["video.codec"]
or video_info["video.pix_fmt"] != reference_video_info["video.pix_fmt"]
):
raise ValueError(
f"Input video {input_path} is not compatible with the reference video {input_video_paths[0]}."
)
# Create a temporary .ffconcat file to list the input video paths
with tempfile.NamedTemporaryFile(mode="w", suffix=".ffconcat", delete=False) as tmp_concatenate_file:
tmp_concatenate_file.write("ffconcat version 1.0\n")
@@ -598,26 +678,20 @@ class _CameraEncoderThread(threading.Thread):
fps: int,
vcodec: str,
pix_fmt: str,
g: int | None,
crf: int | None,
preset: int | None,
codec_options: dict[str, str],
frame_queue: queue.Queue,
result_queue: queue.Queue,
stop_event: threading.Event,
encoder_threads: int | None = None,
):
super().__init__(daemon=True)
self.video_path = video_path
self.fps = fps
self.vcodec = vcodec
self.pix_fmt = pix_fmt
self.g = g
self.crf = crf
self.preset = preset
self.codec_options = codec_options
self.frame_queue = frame_queue
self.result_queue = result_queue
self.stop_event = stop_event
self.encoder_threads = encoder_threads
def run(self) -> None:
from .compute_stats import RunningQuantileStats, auto_downsample_height_width
@@ -653,19 +727,9 @@ class _CameraEncoderThread(threading.Thread):
# Open container on first frame (to get width/height)
if container is None:
height, width = frame_data.shape[:2]
video_options = _get_codec_options(self.vcodec, self.g, self.crf, self.preset)
if self.encoder_threads is not None:
if self.vcodec == "libsvtav1":
lp_param = f"lp={self.encoder_threads}"
if "svtav1-params" in video_options:
video_options["svtav1-params"] += f":{lp_param}"
else:
video_options["svtav1-params"] = lp_param
else:
video_options["threads"] = str(self.encoder_threads)
Path(self.video_path).parent.mkdir(parents=True, exist_ok=True)
container = av.open(str(self.video_path), "w")
output_stream = container.add_stream(self.vcodec, self.fps, options=video_options)
output_stream = container.add_stream(self.vcodec, self.fps, options=self.codec_options)
output_stream.pix_fmt = self.pix_fmt
output_stream.width = width
output_stream.height = height
@@ -731,22 +795,24 @@ class StreamingVideoEncoder:
def __init__(
self,
fps: int,
vcodec: str = "libsvtav1",
pix_fmt: str = "yuv420p",
g: int | None = 2,
crf: int | None = 30,
preset: int | None = None,
camera_encoder: VideoEncoderConfig | None = None,
queue_maxsize: int = 30,
encoder_threads: int | None = None,
):
"""
Args:
fps: Frames per second for the output videos.
camera_encoder: Video encoder settings applied to all cameras.
When ``None``, :func:`camera_encoder_defaults` is used.
encoder_threads: Number of encoder threads (global setting).
``None`` lets the codec decide.
queue_maxsize: Max frames to buffer per camera before
back-pressure drops frames.
"""
self.fps = fps
self.vcodec = resolve_vcodec(vcodec)
self.pix_fmt = pix_fmt
self.g = g
self.crf = crf
self.preset = preset
self._camera_encoder = camera_encoder or camera_encoder_defaults()
self._encoder_threads = encoder_threads
self.queue_maxsize = queue_maxsize
self.encoder_threads = encoder_threads
self._frame_queues: dict[str, queue.Queue] = {}
self._result_queues: dict[str, queue.Queue] = {}
@@ -777,18 +843,17 @@ class StreamingVideoEncoder:
temp_video_dir = Path(tempfile.mkdtemp(dir=temp_dir))
video_path = temp_video_dir / f"{video_key.replace('/', '_')}_streaming.mp4"
vcodec = self._camera_encoder.vcodec
codec_options = self._camera_encoder.get_codec_options(self._encoder_threads, as_strings=True)
encoder_thread = _CameraEncoderThread(
video_path=video_path,
fps=self.fps,
vcodec=self.vcodec,
pix_fmt=self.pix_fmt,
g=self.g,
crf=self.crf,
preset=self.preset,
vcodec=vcodec,
pix_fmt=self._camera_encoder.pix_fmt,
codec_options=codec_options,
frame_queue=frame_queue,
result_queue=result_queue,
stop_event=stop_event,
encoder_threads=self.encoder_threads,
)
encoder_thread.start()
@@ -993,8 +1058,18 @@ def get_audio_info(video_path: Path | str) -> dict:
return audio_info
def get_video_info(video_path: Path | str) -> dict:
# Set logging level
def get_video_info(
video_path: Path | str,
camera_encoder: VideoEncoderConfig | None = None,
) -> dict:
"""Build the ``video.*`` / ``audio.*`` info dict persisted in ``info.json``.
Args:
video_path: Path to the encoded video file to probe.
camera_encoder: If provided, record the exact encoder settings used to encode this
video. Stream-derived values take precedence — encoder fields are only written for keys
not already populated from the video file itself.
"""
logging.getLogger("libav").setLevel(av.logging.WARNING)
# Getting video stream information
@@ -1025,6 +1100,14 @@ def get_video_info(video_path: Path | str) -> dict:
# Adding audio stream information
video_info.update(**get_audio_info(video_path))
# Add additional encoder configuration if provided
if camera_encoder is not None:
for field_name, field_value in asdict(camera_encoder).items():
# vcodec is already populated from the video stream
if field_name == "vcodec":
continue
video_info.setdefault(f"video.{field_name}", field_value)
return video_info
src/lerobot/motors/robstride/robstride.py
+100 -18
View File
@@ -43,6 +43,7 @@ from .tables import (
CAN_CMD_SET_ZERO,
DEFAULT_BAUDRATE,
DEFAULT_TIMEOUT_MS,
HANDSHAKE_TIMEOUT_S,
MODEL_RESOLUTION,
MOTOR_LIMIT_PARAMS,
NORMALIZED_DATA,
@@ -215,14 +216,16 @@ class RobstrideMotorsBus(MotorsBusBase):
self._is_connected = False
raise ConnectionError(f"Failed to connect to CAN bus: {e}") from e
def _query_status_via_clear_fault(self, motor: NameOrID) -> tuple[bool, can.Message | None]:
def _query_status_via_clear_fault(
self, motor: NameOrID, timeout: float = RUNNING_TIMEOUT
) -> tuple[bool, can.Message | None]:
motor_name = self._get_motor_name(motor)
motor_id = self._get_motor_id(motor_name)
recv_id = self._get_motor_recv_id(motor_name)
data = [0xFF] * 7 + [CAN_CMD_CLEAR_FAULT]
msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
self._bus().send(msg)
return self._recv_status_via_clear_fault(expected_recv_id=recv_id)
return self._recv_status_via_clear_fault(expected_recv_id=recv_id, timeout=timeout)
def _recv_status_via_clear_fault(
self, expected_recv_id: int | None = None, timeout: float = RUNNING_TIMEOUT
@@ -280,7 +283,7 @@ class RobstrideMotorsBus(MotorsBusBase):
faulted_motors = []
for motor_name in self.motors:
has_fault, msg = self._query_status_via_clear_fault(motor_name)
has_fault, msg = self._query_status_via_clear_fault(motor_name, timeout=HANDSHAKE_TIMEOUT_S)
if msg is None:
missing_motors.append(motor_name)
elif has_fault:
@@ -505,6 +508,87 @@ class RobstrideMotorsBus(MotorsBusBase):
return responses
def _recv_all_messages_until_quiet(
self,
*,
timeout: float = RUNNING_TIMEOUT,
max_messages: int = 4096,
) -> list[can.Message]:
"""
Receive frames until the bus goes quiet.
Args:
timeout: Poll timeout used for each recv() call. Collection stops
when one recv() times out (quiet gap).
max_messages: Safety cap to prevent unbounded loops.
"""
out: list[can.Message] = []
max_messages = max(1, max_messages)
timeout = max(0.0, timeout)
try:
while len(out) < max_messages:
msg = self._bus().recv(timeout=timeout)
if msg is None:
break
out.append(msg)
except (can.CanError, OSError) as e:
logger.debug(f"Error draining CAN RX queue on {self.port}: {e}")
return out
def _process_feedback_messages(self, messages: list[can.Message]) -> set[int]:
"""
Decode all received feedback frames and update cached motor states.
Returns:
Set of payload recv_ids that were successfully mapped to motors.
"""
processed_recv_ids: set[int] = set()
for msg in messages:
if len(msg.data) < 1:
logger.debug(
f"Dropping short CAN frame on {self.port} "
f"(arb=0x{int(msg.arbitration_id):02X}, data={bytes(msg.data).hex()})"
)
continue
recv_id = int(msg.data[0])
motor_name = self._recv_id_to_motor.get(recv_id)
if motor_name is None:
logger.debug(
f"Unmapped CAN frame on {self.port} "
f"(arb=0x{int(msg.arbitration_id):02X}, recv_id=0x{recv_id:02X}, data={bytes(msg.data).hex()})"
)
continue
self._process_response(motor_name, msg)
processed_recv_ids.add(recv_id)
return processed_recv_ids
def flush_rx_queue(self, poll_timeout_s: float = 0.0005, max_messages: int = 4096) -> int:
"""
Drain pending RX frames from the CAN interface.
This is used by higher-level controllers to drop stale feedback before issuing
a fresh read cycle, so subsequent state reads are based on most recent replies.
It should also be called once when a controller instance is created/connected,
to clear residual frames left on the interface from previous sessions.
"""
drained = 0
poll_timeout_s = max(0.0, poll_timeout_s)
max_messages = max(1, max_messages)
try:
while drained < max_messages:
msg = self._bus().recv(timeout=poll_timeout_s)
if msg is None:
break
drained += 1
except (can.CanError, OSError) as e:
logger.debug(f"Failed to flush CAN RX queue on {self.port}: {e}")
return drained
def _speed_control(
self,
motor: NameOrID,
@@ -644,11 +728,14 @@ class RobstrideMotorsBus(MotorsBusBase):
msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
self._bus().send(msg)
recv_id_to_motor[self._get_motor_recv_id(motor)] = motor_name
# Read every feedback frame until RX goes quiet, then decode all of them.
# This avoids dropping useful frames when responses from different motors interleave.
messages = self._recv_all_messages_until_quiet()
processed_recv_ids = self._process_feedback_messages(messages)
responses = self._recv_all_responses(list(recv_id_to_motor.keys()), timeout=RUNNING_TIMEOUT)
for recv_id, motor_name in recv_id_to_motor.items():
if msg := responses.get(recv_id):
self._process_response(motor_name, msg)
if recv_id not in processed_recv_ids:
logger.warning(f"Packet drop: {motor_name} (ID: 0x{recv_id:02X}). Using last known state.")
def _float_to_uint(self, x: float, x_min: float, x_max: float, bits: int) -> int:
"""Convert float to unsigned integer for CAN transmission."""
@@ -711,7 +798,10 @@ class RobstrideMotorsBus(MotorsBusBase):
try:
self._decode_motor_state(msg.data)
except Exception as e:
logger.warning(f"Failed to decode response from {motor}: {e}")
logger.warning(
f"Failed to decode response from {motor} "
f"(arb=0x{int(msg.arbitration_id):02X}, data={bytes(msg.data).hex()}): {e}"
)
def _get_cached_value(self, motor: str, data_name: str) -> Value:
"""Retrieve a specific value from the state cache."""
@@ -848,20 +938,12 @@ class RobstrideMotorsBus(MotorsBusBase):
self._bus().send(msg)
updated_motors.append(motor)
expected_recv_ids = [self._get_motor_recv_id(motor) for motor in updated_motors]
responses = self._recv_all_responses(expected_recv_ids, timeout=RUNNING_TIMEOUT)
for response in responses.values():
payload_motor_name = self._recv_id_to_motor.get(response.data[0])
if payload_motor_name is not None:
self._process_response(payload_motor_name, response)
else:
# Fallback: still attempt to decode based on payload byte0 mapping.
self._decode_motor_state(response.data)
messages = self._recv_all_messages_until_quiet()
processed_recv_ids = self._process_feedback_messages(messages)
for motor in updated_motors:
recv_id = self._get_motor_recv_id(motor)
if recv_id not in responses:
if recv_id not in processed_recv_ids:
logger.warning(f"Packet drop: {motor} (ID: 0x{recv_id:02X}). Using last known state.")
def read_calibration(self) -> dict[str, MotorCalibration]:
src/lerobot/motors/robstride/tables.py
+2 -1
View File
@@ -114,7 +114,8 @@ CAN_CMD_SAVE_PARAM = 0xAA
CAN_PARAM_ID = 0x7FF
RUNNING_TIMEOUT = 0.001
RUNNING_TIMEOUT = 0.003
HANDSHAKE_TIMEOUT_S = 0.05
PARAM_TIMEOUT = 0.01
STATE_CACHE_TTL_S = 0.02
src/lerobot/policies/eo1/modeling_eo1.py
+3 -2
View File
@@ -28,11 +28,12 @@ import torch.nn.functional as F # noqa: N812
import torch.utils.checkpoint
from torch import Tensor
from lerobot.policies.eo1.configuration_eo1 import EO1Config
from lerobot.policies.pretrained import PreTrainedPolicy
from lerobot.utils.constants import ACTION, OBS_STATE
from lerobot.utils.import_utils import _transformers_available, require_package
from ..pretrained import PreTrainedPolicy
from .configuration_eo1 import EO1Config
if TYPE_CHECKING or _transformers_available:
from transformers.activations import ACT2FN
from transformers.models.qwen2_5_vl import Qwen2_5_VLForConditionalGeneration
src/lerobot/policies/eo1/processor_eo1.py
+2 -1
View File
@@ -22,7 +22,6 @@ from typing import TYPE_CHECKING, Any
import torch
from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
from lerobot.policies.eo1.configuration_eo1 import EO1Config
from lerobot.processor import (
AddBatchDimensionProcessorStep,
ComplementaryDataProcessorStep,
@@ -44,6 +43,8 @@ from lerobot.utils.constants import (
)
from lerobot.utils.import_utils import _transformers_available, require_package
from .configuration_eo1 import EO1Config
if TYPE_CHECKING or _transformers_available:
from transformers.models.qwen2_5_vl import Qwen2_5_VLProcessor
else:
src/lerobot/policies/pi05/modeling_pi05.py
+3 -4
View File
@@ -441,13 +441,13 @@ class PaliGemmaWithExpertModel(
if image.dtype != torch.float32:
image = image.to(torch.float32)
image_outputs = self.paligemma.model.get_image_features(image)
features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
features = image_outputs.pooler_output
if features.dtype != out_dtype:
features = features.to(out_dtype)
return features
def embed_language_tokens(self, tokens: torch.Tensor):
return self.paligemma.model.language_model.embed_tokens(tokens)
return self.paligemma.model.language_model.get_input_embeddings()(tokens)
def forward(
self,
@@ -662,8 +662,7 @@ class PI05Pytorch(nn.Module): # see openpi `PI0Pytorch`
# Process language tokens
def lang_embed_func(tokens):
lang_emb = self.paligemma_with_expert.embed_language_tokens(tokens)
lang_emb_dim = lang_emb.shape[-1]
return lang_emb * math.sqrt(lang_emb_dim)
return lang_emb
lang_emb = self._apply_checkpoint(lang_embed_func, tokens)
embs.append(lang_emb)
src/lerobot/processor/__init__.py
+7
View File
@@ -95,6 +95,13 @@ from .relative_action_processor import (
from .rename_processor import RenameObservationsProcessorStep, rename_stats
from .tokenizer_processor import ActionTokenizerProcessorStep, TokenizerProcessorStep
# RenderMessagesStep is intentionally NOT re-exported here: it pulls in
# `lerobot.datasets.language`, which requires the `[dataset]` extra
# (`datasets`, `pyarrow`). Importing it from the processor package would
# break every base-install consumer of `lerobot.processor`. Users that
# need it import directly:
# from lerobot.processor.render_messages_processor import RenderMessagesStep
__all__ = [
"ActionProcessorStep",
"AddTeleopActionAsComplimentaryDataStep",
src/lerobot/processor/batch_processor.py
+18
View File
@@ -174,6 +174,24 @@ class AddBatchDimensionComplementaryDataStep(ComplementaryDataProcessorStep):
task_index_value = complementary_data["task_index"]
if isinstance(task_index_value, Tensor) and task_index_value.dim() == 0:
complementary_data["task_index"] = task_index_value.unsqueeze(0)
complementary_data.pop("language_persistent", None)
complementary_data.pop("language_events", None)
if "messages" in complementary_data:
messages = complementary_data["messages"]
if isinstance(messages, list) and (not messages or isinstance(messages[0], dict)):
complementary_data["messages"] = [messages]
if "message_streams" in complementary_data:
streams = complementary_data["message_streams"]
if isinstance(streams, list) and (not streams or isinstance(streams[0], str)):
complementary_data["message_streams"] = [streams]
if "target_message_indices" in complementary_data:
indices = complementary_data["target_message_indices"]
if isinstance(indices, list) and (not indices or isinstance(indices[0], int)):
complementary_data["target_message_indices"] = [indices]
return complementary_data
def transform_features(
src/lerobot/processor/converters.py
+20 -16
View File
@@ -153,26 +153,30 @@ def from_tensor_to_numpy(x: torch.Tensor | Any) -> np.ndarray | float | int | An
return x
_COMPLEMENTARY_KEYS = (
"task",
"index",
"task_index",
"episode_index",
"timestamp",
"language_persistent",
"language_events",
"messages",
"message_streams",
"target_message_indices",
)
def _extract_complementary_data(batch: dict[str, Any]) -> dict[str, Any]:
"""
Extract complementary data from a batch dictionary.
"""Extract complementary data from a batch dictionary.
This includes padding flags, task description, and indices.
Args:
batch: The batch dictionary.
Returns:
A dictionary with the extracted complementary data.
Includes padding flags (any key containing ``_is_pad``) plus the fixed
set of metadata / language keys defined in ``_COMPLEMENTARY_KEYS`` —
each only when present in ``batch``.
"""
pad_keys = {k: v for k, v in batch.items() if "_is_pad" in k}
task_key = {"task": batch["task"]} if "task" in batch else {}
subtask_key = {"subtask": batch["subtask"]} if "subtask" in batch else {}
index_key = {"index": batch["index"]} if "index" in batch else {}
task_index_key = {"task_index": batch["task_index"]} if "task_index" in batch else {}
episode_index_key = {"episode_index": batch["episode_index"]} if "episode_index" in batch else {}
return {**pad_keys, **task_key, **subtask_key, **index_key, **task_index_key, **episode_index_key}
extras = {k: batch[k] for k in _COMPLEMENTARY_KEYS if k in batch}
return {**pad_keys, **extras}
def create_transition(
src/lerobot/processor/render_messages_processor.py
+84
View File
@@ -0,0 +1,84 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations
from dataclasses import dataclass
from typing import Any
from lerobot.configs import PipelineFeatureType, PolicyFeature
from lerobot.configs.recipe import TrainingRecipe
from lerobot.datasets.language import LANGUAGE_EVENTS, LANGUAGE_PERSISTENT
from lerobot.datasets.language_render import render_sample
from lerobot.types import EnvTransition, TransitionKey
from lerobot.utils.utils import unwrap_scalar
from .pipeline import ProcessorStep, ProcessorStepRegistry
@dataclass
@ProcessorStepRegistry.register(name="render_messages_processor")
class RenderMessagesStep(ProcessorStep):
"""Processor step that turns raw language columns into rendered chat messages.
Reads ``language_persistent`` and ``language_events`` from the transition's
complementary data, renders them through ``recipe`` at the sample timestamp,
and replaces the raw columns with the resulting ``messages`` /
``message_streams`` / ``target_message_indices`` keys.
"""
recipe: TrainingRecipe
dataset_ctx: Any | None = None
def __call__(self, transition: EnvTransition) -> EnvTransition | None:
"""Render messages for a single transition; return ``None`` to drop it."""
complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
persistent = complementary_data.get(LANGUAGE_PERSISTENT) or []
events = complementary_data.get(LANGUAGE_EVENTS) or []
if not persistent and not events:
return transition
timestamp = complementary_data.get("timestamp")
if timestamp is None:
raise KeyError("RenderMessagesStep requires sample timestamp in complementary data.")
sample_idx = complementary_data.get("index", 0)
rendered = render_sample(
recipe=self.recipe,
persistent=persistent,
events=events,
t=unwrap_scalar(timestamp),
sample_idx=int(unwrap_scalar(sample_idx)),
task=complementary_data.get("task"),
dataset_ctx=self.dataset_ctx,
)
if rendered is None:
return None
new_transition = transition.copy()
new_complementary_data = dict(complementary_data)
new_complementary_data.pop(LANGUAGE_PERSISTENT, None)
new_complementary_data.pop(LANGUAGE_EVENTS, None)
new_complementary_data.update(rendered)
new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
return new_transition
def transform_features(
self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
"""Pass features through unchanged; rendering only touches complementary data."""
return features
src/lerobot/rewards/classifier/modeling_classifier.py
+3 -2
View File
@@ -17,10 +17,11 @@ import logging
import torch
from torch import Tensor, nn
from lerobot.rewards.classifier.configuration_classifier import RewardClassifierConfig
from lerobot.rewards.pretrained import PreTrainedRewardModel
from lerobot.utils.constants import OBS_IMAGE, REWARD
from ..pretrained import PreTrainedRewardModel
from .configuration_classifier import RewardClassifierConfig
class ClassifierOutput:
"""Wrapper for classifier outputs with additional metadata."""
src/lerobot/rewards/classifier/processor_classifier.py
+2 -1
View File
@@ -25,7 +25,8 @@ from lerobot.processor import (
policy_action_to_transition,
transition_to_policy_action,
)
from lerobot.rewards.classifier.configuration_classifier import RewardClassifierConfig
from .configuration_classifier import RewardClassifierConfig
def make_classifier_processor(
src/lerobot/rewards/factory.py
+4 -3
View File
@@ -22,9 +22,10 @@ import torch
from lerobot.configs.rewards import RewardModelConfig
from lerobot.processor import PolicyAction, PolicyProcessorPipeline
from lerobot.rewards.classifier.configuration_classifier import RewardClassifierConfig
from lerobot.rewards.pretrained import PreTrainedRewardModel
from lerobot.rewards.sarm.configuration_sarm import SARMConfig
from .classifier.configuration_classifier import RewardClassifierConfig
from .pretrained import PreTrainedRewardModel
from .sarm.configuration_sarm import SARMConfig
def get_reward_model_class(name: str) -> type[PreTrainedRewardModel]:
src/lerobot/rewards/sarm/compute_rabc_weights.py
+4 -3
View File
@@ -58,9 +58,10 @@ import torch
from tqdm import tqdm
from lerobot.datasets import LeRobotDataset
from lerobot.rewards.sarm.modeling_sarm import SARMRewardModel
from lerobot.rewards.sarm.processor_sarm import make_sarm_pre_post_processors
from lerobot.rewards.sarm.sarm_utils import normalize_stage_tau
from .modeling_sarm import SARMRewardModel
from .processor_sarm import make_sarm_pre_post_processors
from .sarm_utils import normalize_stage_tau
def get_reward_model_path_from_parquet(parquet_path: Path) -> str | None:
src/lerobot/rewards/sarm/modeling_sarm.py
+5 -4
View File
@@ -32,13 +32,14 @@ import torch.nn as nn
import torch.nn.functional as F # noqa: N812
from torch import Tensor
from lerobot.rewards.pretrained import PreTrainedRewardModel
from lerobot.rewards.sarm.configuration_sarm import SARMConfig
from lerobot.rewards.sarm.sarm_utils import (
from lerobot.utils.constants import OBS_STR
from ..pretrained import PreTrainedRewardModel
from .configuration_sarm import SARMConfig
from .sarm_utils import (
normalize_stage_tau,
pad_state_to_max_dim,
)
from lerobot.utils.constants import OBS_STR
class StageTransformer(nn.Module):
src/lerobot/rewards/sarm/processor_sarm.py
+5 -4
View File
@@ -58,15 +58,16 @@ from lerobot.processor import (
policy_action_to_transition,
transition_to_policy_action,
)
from lerobot.rewards.sarm.configuration_sarm import SARMConfig
from lerobot.rewards.sarm.sarm_utils import (
from lerobot.types import EnvTransition, PolicyAction, TransitionKey
from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
from .configuration_sarm import SARMConfig
from .sarm_utils import (
apply_rewind_augmentation,
compute_absolute_indices,
find_stage_and_tau,
pad_state_to_max_dim,
)
from lerobot.types import EnvTransition, PolicyAction, TransitionKey
from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
class SARMEncodingProcessorStep(ProcessorStep):
src/lerobot/robots/bi_rebot_b601_follower/__init__.py
+20
View File
@@ -0,0 +1,20 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .bi_rebot_b601_follower import BiRebotB601Follower
from .config_bi_rebot_b601_follower import BiRebotB601FollowerConfig
__all__ = ["BiRebotB601Follower", "BiRebotB601FollowerConfig"]
src/lerobot/robots/bi_rebot_b601_follower/bi_rebot_b601_follower.py
+150
View File
@@ -0,0 +1,150 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
from functools import cached_property
from lerobot.types import RobotAction, RobotObservation
from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
from ..rebot_b601_follower import RebotB601Follower, RebotB601FollowerRobotConfig
from ..robot import Robot
from .config_bi_rebot_b601_follower import BiRebotB601FollowerConfig
logger = logging.getLogger(__name__)
class BiRebotB601Follower(Robot):
"""Bimanual Seeed Studio reBot B601-DM follower.
Composes two single-arm :class:`RebotB601Follower` instances. Observation and
action keys of each arm are namespaced with a ``left_`` / ``right_`` prefix.
"""
config_class = BiRebotB601FollowerConfig
name = "bi_rebot_b601_follower"
def __init__(self, config: BiRebotB601FollowerConfig):
super().__init__(config)
self.config = config
left_arm_config = RebotB601FollowerRobotConfig(
id=f"{config.id}_left" if config.id else None,
calibration_dir=config.calibration_dir,
port=config.left_arm_config.port,
can_adapter=config.left_arm_config.can_adapter,
dm_serial_baud=config.left_arm_config.dm_serial_baud,
disable_torque_on_disconnect=config.left_arm_config.disable_torque_on_disconnect,
max_relative_target=config.left_arm_config.max_relative_target,
cameras=config.left_arm_config.cameras,
motor_can_ids=config.left_arm_config.motor_can_ids,
pos_vel_velocity=config.left_arm_config.pos_vel_velocity,
gripper_torque_ratio=config.left_arm_config.gripper_torque_ratio,
joint_limits=config.left_arm_config.joint_limits,
)
right_arm_config = RebotB601FollowerRobotConfig(
id=f"{config.id}_right" if config.id else None,
calibration_dir=config.calibration_dir,
port=config.right_arm_config.port,
can_adapter=config.right_arm_config.can_adapter,
dm_serial_baud=config.right_arm_config.dm_serial_baud,
disable_torque_on_disconnect=config.right_arm_config.disable_torque_on_disconnect,
max_relative_target=config.right_arm_config.max_relative_target,
cameras=config.right_arm_config.cameras,
motor_can_ids=config.right_arm_config.motor_can_ids,
pos_vel_velocity=config.right_arm_config.pos_vel_velocity,
gripper_torque_ratio=config.right_arm_config.gripper_torque_ratio,
joint_limits=config.right_arm_config.joint_limits,
)
self.left_arm = RebotB601Follower(left_arm_config)
self.right_arm = RebotB601Follower(right_arm_config)
# Only for compatibility with parts of the codebase that expect `robot.cameras`.
self.cameras = {**self.left_arm.cameras, **self.right_arm.cameras}
@property
def _motors_ft(self) -> dict[str, type]:
return {
**{f"left_{k}": v for k, v in self.left_arm._motors_ft.items()},
**{f"right_{k}": v for k, v in self.right_arm._motors_ft.items()},
}
@property
def _cameras_ft(self) -> dict[str, tuple]:
return {
**{f"left_{k}": v for k, v in self.left_arm._cameras_ft.items()},
**{f"right_{k}": v for k, v in self.right_arm._cameras_ft.items()},
}
@cached_property
def observation_features(self) -> dict[str, type | tuple]:
return {**self._motors_ft, **self._cameras_ft}
@cached_property
def action_features(self) -> dict[str, type]:
return self._motors_ft
@property
def is_connected(self) -> bool:
return self.left_arm.is_connected and self.right_arm.is_connected
@check_if_already_connected
def connect(self, calibrate: bool = True) -> None:
self.left_arm.connect(calibrate)
self.right_arm.connect(calibrate)
@property
def is_calibrated(self) -> bool:
return self.left_arm.is_calibrated and self.right_arm.is_calibrated
def calibrate(self) -> None:
self.left_arm.calibrate()
self.right_arm.calibrate()
def configure(self) -> None:
self.left_arm.configure()
self.right_arm.configure()
@check_if_not_connected
def get_observation(self) -> RobotObservation:
obs_dict = {}
obs_dict.update({f"left_{k}": v for k, v in self.left_arm.get_observation().items()})
obs_dict.update({f"right_{k}": v for k, v in self.right_arm.get_observation().items()})
return obs_dict
@check_if_not_connected
def send_action(self, action: RobotAction) -> RobotAction:
left_action = {
key.removeprefix("left_"): value for key, value in action.items() if key.startswith("left_")
}
right_action = {
key.removeprefix("right_"): value for key, value in action.items() if key.startswith("right_")
}
sent_action_left = self.left_arm.send_action(left_action)
sent_action_right = self.right_arm.send_action(right_action)
return {
**{f"left_{k}": v for k, v in sent_action_left.items()},
**{f"right_{k}": v for k, v in sent_action_right.items()},
}
@check_if_not_connected
def disconnect(self) -> None:
self.left_arm.disconnect()
self.right_arm.disconnect()
src/lerobot/robots/bi_rebot_b601_follower/config_bi_rebot_b601_follower.py
+29
View File
@@ -0,0 +1,29 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass
from ..config import RobotConfig
from ..rebot_b601_follower import RebotB601FollowerConfig
@RobotConfig.register_subclass("bi_rebot_b601_follower")
@dataclass
class BiRebotB601FollowerConfig(RobotConfig):
"""Configuration class for the bimanual reBot B601-DM follower robot."""
left_arm_config: RebotB601FollowerConfig
right_arm_config: RebotB601FollowerConfig
src/lerobot/robots/rebot_b601_follower/__init__.py
+20
View File
@@ -0,0 +1,20 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .config_rebot_b601_follower import RebotB601FollowerConfig, RebotB601FollowerRobotConfig
from .rebot_b601_follower import RebotB601Follower
__all__ = ["RebotB601Follower", "RebotB601FollowerConfig", "RebotB601FollowerRobotConfig"]
src/lerobot/robots/rebot_b601_follower/config_rebot_b601_follower.py
+94
View File
@@ -0,0 +1,94 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass, field
from lerobot.cameras import CameraConfig
from ..config import RobotConfig
@dataclass
class RebotB601FollowerConfig:
"""Base configuration class for the Seeed Studio reBot B601-DM follower arm.
The B601-DM is a 6-DOF arm plus gripper driven by Damiao CAN motors. Motor
communication goes through the ``motorbridge`` package.
"""
# Communication port. For ``can_adapter="damiao"`` this is the Damiao serial
# bridge device (e.g. "/dev/ttyACM0"); for ``can_adapter="socketcan"`` it is
# the CAN channel name (e.g. "can0").
port: str
# CAN adapter type:
# "damiao" - Damiao dedicated serial bridge (default)
# "socketcan" - SocketCAN based adapters (PCAN, slcan, embedded controllers, ...)
can_adapter: str = "damiao"
# Baud rate for the Damiao serial bridge (only used when can_adapter="damiao").
dm_serial_baud: int = 921600
disable_torque_on_disconnect: bool = True
# `max_relative_target` limits the magnitude of the relative positional target
# vector for safety purposes (in degrees). Set to a positive scalar to apply the
# same value to all motors, or to a dict mapping motor names to per-motor values.
max_relative_target: float | dict[str, float] | None = None
# cameras
cameras: dict[str, CameraConfig] = field(default_factory=dict)
# Maps motor names to their (send_can_id, recv_can_id) pair.
motor_can_ids: dict[str, tuple[int, int]] = field(
default_factory=lambda: {
"shoulder_pan": (0x01, 0x11),
"shoulder_lift": (0x02, 0x12),
"elbow_flex": (0x03, 0x13),
"wrist_flex": (0x04, 0x14),
"wrist_yaw": (0x05, 0x15),
"wrist_roll": (0x06, 0x16),
"gripper": (0x07, 0x17),
}
)
# Target velocity for joints running in POS_VEL mode, in degrees/s. A scalar is
# applied to every joint; a list provides one value per joint (in motor order).
pos_vel_velocity: float | list[float] = field(default_factory=lambda: [150.0] * 7)
# Torque/current ratio for the gripper's FORCE_POS mode, in range [0, 1].
gripper_torque_ratio: float = 0.1
# Soft joint limits (degrees). These are clipped against on every action.
joint_limits: dict[str, tuple[float, float]] = field(
default_factory=lambda: {
"shoulder_pan": (-145.0, 145.0),
"shoulder_lift": (-170.0, 1.0),
"elbow_flex": (-200.0, 1.0),
"wrist_flex": (-80.0, 90.0),
"wrist_yaw": (-90.0, 90.0),
"wrist_roll": (-90.0, 90.0),
"gripper": (-270.0, 0.0),
}
)
@RobotConfig.register_subclass("rebot_b601_follower")
@dataclass
class RebotB601FollowerRobotConfig(RobotConfig, RebotB601FollowerConfig):
"""Registered configuration for the reBot B601-DM follower robot."""
pass
src/lerobot/robots/rebot_b601_follower/rebot_b601_follower.py
+289
View File
@@ -0,0 +1,289 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
import math
import time
from functools import cached_property
from typing import TYPE_CHECKING
from lerobot.cameras import make_cameras_from_configs
from lerobot.motors import MotorCalibration
from lerobot.types import RobotAction, RobotObservation
from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
from lerobot.utils.import_utils import _motorbridge_available, require_package
from ..robot import Robot
from ..utils import ensure_safe_goal_position
from .config_rebot_b601_follower import RebotB601FollowerRobotConfig
if TYPE_CHECKING or _motorbridge_available:
from motorbridge import Controller as MotorBridgeController, Mode as MotorBridgeMode
else:
MotorBridgeController = None
MotorBridgeMode = None
logger = logging.getLogger(__name__)
# Joint controlled in FORCE_POS mode; every other joint runs in POS_VEL mode.
GRIPPER_MOTOR = "gripper"
# Per-joint Damiao motor models for the B601-DM (passed to motorbridge).
MOTOR_MODELS = {
"shoulder_pan": "4340P",
"shoulder_lift": "4340P",
"elbow_flex": "4340P",
"wrist_flex": "4310",
"wrist_yaw": "4310",
"wrist_roll": "4310",
"gripper": "4310",
}
_ENSURE_MODE_RETRIES = 9
_SETTLE_SEC = 0.01
_ZERO_SETTLE_SEC = 0.1
class RebotB601Follower(Robot):
"""Seeed Studio reBot B601-DM follower arm (6-DOF + gripper, Damiao CAN motors).
Motor communication is handled by the ``motorbridge`` package over a CAN bus,
reached either through a Damiao serial bridge or a SocketCAN adapter.
"""
config_class = RebotB601FollowerRobotConfig
name = "rebot_b601_follower"
def __init__(self, config: RebotB601FollowerRobotConfig):
require_package("motorbridge", extra="rebot")
super().__init__(config)
self.config = config
self.bus: MotorBridgeController | None = None
self.motors: dict = {}
self.motor_names = list(config.motor_can_ids.keys())
self.cameras = make_cameras_from_configs(config.cameras)
@property
def _motors_ft(self) -> dict[str, type]:
return {f"{motor}.pos": float for motor in self.motor_names}
@property
def _cameras_ft(self) -> dict[str, tuple]:
return {
cam: (self.config.cameras[cam].height, self.config.cameras[cam].width, 3) for cam in self.cameras
}
@cached_property
def observation_features(self) -> dict[str, type | tuple]:
return {**self._motors_ft, **self._cameras_ft}
@cached_property
def action_features(self) -> dict[str, type]:
return self._motors_ft
@property
def is_connected(self) -> bool:
return self.bus is not None and all(cam.is_connected for cam in self.cameras.values())
@check_if_already_connected
def connect(self, calibrate: bool = True) -> None:
logger.info(f"Connecting {self} on {self.config.port} (adapter={self.config.can_adapter})...")
if self.config.can_adapter == "damiao":
self.bus = MotorBridgeController.from_dm_serial(
serial_port=self.config.port,
baud=self.config.dm_serial_baud,
)
elif self.config.can_adapter == "socketcan":
self.bus = MotorBridgeController(channel=self.config.port)
else:
raise ValueError(
f"Unsupported can_adapter '{self.config.can_adapter}'. Use 'damiao' or 'socketcan'."
)
for motor_name, (send_id, recv_id) in self.config.motor_can_ids.items():
self.motors[motor_name] = self.bus.add_damiao_motor(send_id, recv_id, MOTOR_MODELS[motor_name])
if not self.is_calibrated and calibrate:
logger.info(
"Mismatch between calibration values in the motor and the calibration file or no calibration file found"
)
self.calibrate()
for cam in self.cameras.values():
cam.connect()
self.configure()
logger.info(f"{self} connected.")
@property
def is_calibrated(self) -> bool:
return bool(self.calibration)
def calibrate(self) -> None:
if self.calibration:
user_input = input(
f"Press ENTER to use provided calibration file associated with the id {self.id}, "
"or type 'c' and press ENTER to run calibration: "
)
if user_input.strip().lower() != "c":
logger.info(f"Using calibration file associated with the id {self.id}")
return
logger.info(f"\nRunning calibration of {self}")
self.bus.disable_all()
print(
"\nCalibration: set zero position.\n"
"Manually move the reBot B601 to its ZERO POSITION and close the gripper.\n"
"See the B601 manual for the zero pose (the default sit-down position).\n"
)
input("Press ENTER when ready...")
for motor in self.motors.values():
motor.set_zero_position()
time.sleep(_ZERO_SETTLE_SEC)
logger.info("Arm zero position set.")
self.calibration = {}
for motor_name, (send_id, _recv_id) in self.config.motor_can_ids.items():
range_min, range_max = self.config.joint_limits[motor_name]
self.calibration[motor_name] = MotorCalibration(
id=send_id,
drive_mode=0,
homing_offset=0,
range_min=int(range_min),
range_max=int(range_max),
)
self._save_calibration()
print(f"Calibration saved to {self.calibration_fpath}")
def configure(self) -> None:
self.bus.enable_all()
for motor_name, motor in self.motors.items():
target_mode = (
MotorBridgeMode.FORCE_POS if motor_name == GRIPPER_MOTOR else MotorBridgeMode.POS_VEL
)
for attempt in range(_ENSURE_MODE_RETRIES + 1):
try:
motor.ensure_mode(target_mode)
break
except Exception:
if attempt == _ENSURE_MODE_RETRIES:
raise
time.sleep(_SETTLE_SEC)
logger.debug(f"{motor_name} mode set to {target_mode}")
@check_if_not_connected
def disable_torque(self) -> None:
"""Disable motor torque so the arm can be moved by hand (read-only debugging)."""
self.bus.disable_all()
logger.info(f"{self} torque disabled.")
def _present_pos(self) -> dict[str, float]:
"""Read present joint positions in degrees."""
for motor in self.motors.values():
motor.request_feedback()
try:
self.bus.poll_feedback_once()
except Exception:
logger.warning("CAN bus poll feedback failed.")
present_pos = {}
for motor_name, motor in self.motors.items():
state = motor.get_state()
present_pos[motor_name] = math.degrees(state.pos) if state is not None else 0.0
return present_pos
@check_if_not_connected
def get_observation(self) -> RobotObservation:
start = time.perf_counter()
obs_dict = {f"{motor}.pos": pos for motor, pos in self._present_pos().items()}
dt_ms = (time.perf_counter() - start) * 1e3
logger.debug(f"{self} read state: {dt_ms:.1f}ms")
for cam_key, cam in self.cameras.items():
start = time.perf_counter()
obs_dict[cam_key] = cam.read_latest()
dt_ms = (time.perf_counter() - start) * 1e3
logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
return obs_dict
@check_if_not_connected
def send_action(self, action: RobotAction) -> RobotAction:
"""Command the arm to a target joint configuration.
Positions are expressed in degrees. The relative action magnitude may be
clipped depending on `max_relative_target`, so the action actually sent is
always returned.
"""
goal_pos = {key.removesuffix(".pos"): val for key, val in action.items() if key.endswith(".pos")}
# Clip against soft joint limits.
for motor_name in list(goal_pos):
if motor_name in self.config.joint_limits:
min_limit, max_limit = self.config.joint_limits[motor_name]
clipped = max(min_limit, min(max_limit, goal_pos[motor_name]))
if clipped != goal_pos[motor_name]:
logger.debug(f"Clipped {motor_name} from {goal_pos[motor_name]:.2f} to {clipped:.2f}")
goal_pos[motor_name] = clipped
# Tolerate 6-DOF leaders that have no wrist_yaw joint by holding it at zero.
# This is intentional: it lets a 6-DOF leader such as the SO-100 / SO-101
# (so100_leader / so101_leader) teleoperate this 7-DOF follower — the missing
# wrist_yaw command is simply treated as 0.0 instead of raising.
if "wrist_yaw" not in goal_pos:
goal_pos["wrist_yaw"] = 0.0
# Cap relative target when too far from the present position.
if self.config.max_relative_target is not None:
present_pos = self._present_pos()
goal_present_pos = {key: (g, present_pos.get(key, g)) for key, g in goal_pos.items()}
goal_pos = ensure_safe_goal_position(goal_present_pos, self.config.max_relative_target)
for motor_name, position_deg in goal_pos.items():
motor = self.motors.get(motor_name)
if motor is None:
continue
idx = self.motor_names.index(motor_name)
vel_deg_s = (
self.config.pos_vel_velocity[idx]
if isinstance(self.config.pos_vel_velocity, list)
else self.config.pos_vel_velocity
)
pos_rad = math.radians(position_deg)
vel_rad = math.radians(vel_deg_s)
if motor_name == GRIPPER_MOTOR:
motor.send_force_pos(pos_rad, vel_rad, self.config.gripper_torque_ratio)
else:
motor.send_pos_vel(pos_rad, vel_rad)
return {f"{motor}.pos": val for motor, val in goal_pos.items()}
@check_if_not_connected
def disconnect(self) -> None:
for motor in self.motors.values():
if self.config.disable_torque_on_disconnect:
motor.disable()
motor.clear_error()
motor.close()
self.bus.close()
self.bus = None
self.motors = {}
for cam in self.cameras.values():
cam.disconnect()
logger.info(f"{self} disconnected.")
src/lerobot/robots/utils.py
+8
View File
@@ -68,6 +68,14 @@ def make_robot_from_config(config: RobotConfig) -> Robot:
from .bi_openarm_follower import BiOpenArmFollower
return BiOpenArmFollower(config)
elif config.type == "rebot_b601_follower":
from .rebot_b601_follower import RebotB601Follower
return RebotB601Follower(config)
elif config.type == "bi_rebot_b601_follower":
from .bi_rebot_b601_follower import BiRebotB601Follower
return BiRebotB601Follower(config)
elif config.type == "mock_robot":
from tests.mocks.mock_robot import MockRobot
src/lerobot/rollout/context.py
+2 -2
View File
@@ -332,7 +332,7 @@ def build_rollout_context(
cfg.dataset.repo_id,
root=cfg.dataset.root,
batch_encoding_size=cfg.dataset.video_encoding_batch_size,
vcodec=cfg.dataset.vcodec,
camera_encoder=cfg.dataset.camera_encoder,
streaming_encoding=cfg.dataset.streaming_encoding,
encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
encoder_threads=cfg.dataset.encoder_threads,
@@ -367,7 +367,7 @@ def build_rollout_context(
image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera
* len(robot.cameras if hasattr(robot, "cameras") else []),
batch_encoding_size=cfg.dataset.video_encoding_batch_size,
vcodec=cfg.dataset.vcodec,
camera_encoder=cfg.dataset.camera_encoder,
streaming_encoding=cfg.dataset.streaming_encoding,
encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
encoder_threads=cfg.dataset.encoder_threads,
src/lerobot/scripts/lerobot_calibrate.py
+4
View File
@@ -39,6 +39,7 @@ from lerobot.robots import ( # noqa: F401
Robot,
RobotConfig,
bi_openarm_follower,
bi_rebot_b601_follower,
bi_so_follower,
hope_jr,
koch_follower,
@@ -46,12 +47,14 @@ from lerobot.robots import ( # noqa: F401
make_robot_from_config,
omx_follower,
openarm_follower,
rebot_b601_follower,
so_follower,
)
from lerobot.teleoperators import ( # noqa: F401
Teleoperator,
TeleoperatorConfig,
bi_openarm_leader,
bi_rebot_102_leader,
bi_so_leader,
homunculus,
koch_leader,
@@ -59,6 +62,7 @@ from lerobot.teleoperators import ( # noqa: F401
omx_leader,
openarm_leader,
openarm_mini,
rebot_102_leader,
so_leader,
unitree_g1,
)
src/lerobot/scripts/lerobot_edit_dataset.py
+93 -12
View File
@@ -178,6 +178,31 @@ Recompute stats for relative actions and push to hub:
--operation.num_workers 4 \
--push_to_hub true
Re-encode all videos in a dataset (saves to lerobot/pusht_reencoded by default):
lerobot-edit-dataset \
--repo_id lerobot/pusht \
--operation.type reencode_videos \
--operation.camera_encoder.vcodec h264 \
--operation.camera_encoder.pix_fmt yuv420p \
--operation.camera_encoder.crf 23
Re-encode videos into a new dataset using 4 parallel processes:
lerobot-edit-dataset \
--repo_id lerobot/pusht \
--new_repo_id lerobot/pusht_h264 \
--operation.type reencode_videos \
--operation.camera_encoder.vcodec h264 \
--operation.camera_encoder.crf 23 \
--operation.num_workers 4
Re-encode videos in-place (overwrites original dataset):
lerobot-edit-dataset \
--repo_id lerobot/pusht \
--new_repo_id lerobot/pusht \
--operation.type reencode_videos \
--operation.camera_encoder.vcodec h264 \
--operation.overwrite true
Using JSON config file:
lerobot-edit-dataset \
--config_path path/to/edit_config.json
@@ -187,12 +212,12 @@ import abc
import logging
import shutil
import sys
from dataclasses import dataclass
from dataclasses import dataclass, field
from pathlib import Path
import draccus
from lerobot.configs import parser
from lerobot.configs import VideoEncoderConfig, camera_encoder_defaults, parser
from lerobot.datasets import (
LeRobotDataset,
convert_image_to_video_dataset,
@@ -200,6 +225,7 @@ from lerobot.datasets import (
merge_datasets,
modify_tasks,
recompute_stats,
reencode_dataset,
remove_feature,
split_dataset,
)
@@ -250,11 +276,7 @@ class ModifyTasksConfig(OperationConfig):
@dataclass
class ConvertImageToVideoConfig(OperationConfig):
output_dir: str | None = None
vcodec: str = "libsvtav1"
pix_fmt: str = "yuv420p"
g: int = 2
crf: int = 30
fast_decode: int = 0
camera_encoder: VideoEncoderConfig = field(default_factory=camera_encoder_defaults)
episode_indices: list[int] | None = None
num_workers: int = 4
max_episodes_per_batch: int | None = None
@@ -272,6 +294,15 @@ class RecomputeStatsConfig(OperationConfig):
overwrite: bool = False
@OperationConfig.register_subclass("reencode_videos")
@dataclass
class ReencodeVideosConfig(OperationConfig):
camera_encoder: VideoEncoderConfig = field(default_factory=camera_encoder_defaults)
num_workers: int = 0
encoder_threads: int | None = None
overwrite: bool = False
@OperationConfig.register_subclass("info")
@dataclass
class InfoConfig(OperationConfig):
@@ -557,11 +588,7 @@ def handle_convert_image_to_video(cfg: EditDatasetConfig) -> None:
dataset=dataset,
output_dir=output_dir,
repo_id=output_repo_id,
vcodec=getattr(cfg.operation, "vcodec", "libsvtav1"),
pix_fmt=getattr(cfg.operation, "pix_fmt", "yuv420p"),
g=getattr(cfg.operation, "g", 2),
crf=getattr(cfg.operation, "crf", 30),
fast_decode=getattr(cfg.operation, "fast_decode", 0),
camera_encoder=getattr(cfg.operation, "camera_encoder", None) or camera_encoder_defaults(),
episode_indices=getattr(cfg.operation, "episode_indices", None),
num_workers=getattr(cfg.operation, "num_workers", 4),
max_episodes_per_batch=getattr(cfg.operation, "max_episodes_per_batch", None),
@@ -642,6 +669,58 @@ def handle_recompute_stats(cfg: EditDatasetConfig) -> None:
dataset.push_to_hub()
def handle_reencode_videos(cfg: EditDatasetConfig) -> None:
if not isinstance(cfg.operation, ReencodeVideosConfig):
raise ValueError("Operation config must be ReencodeVideosConfig")
output_repo_id, input_root, output_root = _resolve_io_paths(
cfg.repo_id,
cfg.new_repo_id,
cfg.root,
cfg.new_root,
default_new_repo_id=f"{cfg.repo_id}_reencoded",
)
in_place = output_root == input_root
if in_place and not cfg.operation.overwrite:
raise ValueError(
f"reencode_videos would overwrite the dataset in-place at {input_root}. "
"Pass --operation.overwrite true to allow in-place modification, "
"or use --new_repo_id / --new_root to write to a different location. "
f"Default output repo_id when neither is set: '{cfg.repo_id}_reencoded'."
)
if in_place:
logging.warning(
f"Overwriting dataset videos in-place at {input_root}. The original videos will be lost."
)
dataset = LeRobotDataset(cfg.repo_id, root=input_root)
else:
logging.info(f"Copying dataset from {input_root} to {output_root}")
if output_root.exists():
backup_path = output_root.with_name(output_root.name + "_old")
logging.warning(f"Output directory {output_root} already exists. Moving to {backup_path}")
if backup_path.exists():
shutil.rmtree(backup_path)
shutil.move(output_root, backup_path)
shutil.copytree(input_root, output_root)
dataset = LeRobotDataset(output_repo_id, root=output_root)
logging.info(f"Re-encoding videos in {output_repo_id} with {cfg.operation.camera_encoder}")
reencode_dataset(
dataset,
camera_encoder=cfg.operation.camera_encoder,
encoder_threads=cfg.operation.encoder_threads,
num_workers=cfg.operation.num_workers,
)
logging.info(f"All videos re-encoded at {dataset.root}")
if cfg.push_to_hub:
logging.info(f"Pushing to hub as {output_repo_id}...")
dataset.push_to_hub()
def _get_dataset_size(repo_path):
import os
@@ -715,6 +794,8 @@ def edit_dataset(cfg: EditDatasetConfig) -> None:
handle_convert_image_to_video(cfg)
elif operation_type == "recompute_stats":
handle_recompute_stats(cfg)
elif operation_type == "reencode_videos":
handle_reencode_videos(cfg)
elif operation_type == "info":
handle_info(cfg)
else:
src/lerobot/scripts/lerobot_find_joint_limits.py
+4
View File
@@ -45,16 +45,19 @@ from lerobot.model import RobotKinematics
from lerobot.robots import ( # noqa: F401
RobotConfig,
bi_openarm_follower,
bi_rebot_b601_follower,
bi_so_follower,
koch_follower,
make_robot_from_config,
omx_follower,
openarm_follower,
rebot_b601_follower,
so_follower,
)
from lerobot.teleoperators import ( # noqa: F401
TeleoperatorConfig,
bi_openarm_leader,
bi_rebot_102_leader,
bi_so_leader,
gamepad,
koch_leader,
@@ -62,6 +65,7 @@ from lerobot.teleoperators import ( # noqa: F401
omx_leader,
openarm_leader,
openarm_mini,
rebot_102_leader,
so_leader,
)
from lerobot.utils.robot_utils import precise_sleep
src/lerobot/scripts/lerobot_record.py
+30 -5
View File
@@ -63,6 +63,27 @@ lerobot-record \\
--dataset.streaming_encoding=true \\
--dataset.encoder_threads=2
```
Example recording with custom video encoding parameters:
```shell
lerobot-record \\
--robot.type=so100_follower \\
--robot.port=/dev/tty.usbmodem58760431541 \\
--robot.cameras="{laptop: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \\
--robot.id=black \\
--teleop.type=so100_leader \\
--teleop.port=/dev/tty.usbmodem58760431551 \\
--teleop.id=blue \\
--dataset.repo_id=<my_username>/<my_dataset_name> \\
--dataset.num_episodes=2 \\
--dataset.single_task="Grab the cube" \\
--dataset.streaming_encoding=true \\
--dataset.encoder_threads=2 \\
--dataset.camera_encoder.vcodec=h264 \\
--dataset.camera_encoder.preset=fast \\
--dataset.camera_encoder.extra_options={"tune": "film", "profile:v": "high", "bf": 2} \\
--display_data=true
```
"""
import logging
@@ -99,6 +120,7 @@ from lerobot.robots import ( # noqa: F401
Robot,
RobotConfig,
bi_openarm_follower,
bi_rebot_b601_follower,
bi_so_follower,
earthrover_mini_plus,
hope_jr,
@@ -107,6 +129,7 @@ from lerobot.robots import ( # noqa: F401
omx_follower,
openarm_follower,
reachy2,
rebot_b601_follower,
so_follower,
unitree_g1 as unitree_g1_robot,
)
@@ -114,6 +137,7 @@ from lerobot.teleoperators import ( # noqa: F401
Teleoperator,
TeleoperatorConfig,
bi_openarm_leader,
bi_rebot_102_leader,
bi_so_leader,
homunculus,
koch_leader,
@@ -122,6 +146,7 @@ from lerobot.teleoperators import ( # noqa: F401
openarm_leader,
openarm_mini,
reachy2_teleoperator,
rebot_102_leader,
so_leader,
unitree_g1,
)
@@ -377,10 +402,10 @@ def record(
cfg.dataset.repo_id,
root=cfg.dataset.root,
batch_encoding_size=cfg.dataset.video_encoding_batch_size,
vcodec=cfg.dataset.vcodec,
camera_encoder=cfg.dataset.camera_encoder,
encoder_threads=cfg.dataset.encoder_threads,
streaming_encoding=cfg.dataset.streaming_encoding,
encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
encoder_threads=cfg.dataset.encoder_threads,
image_writer_processes=cfg.dataset.num_image_writer_processes if num_cameras > 0 else 0,
image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera * num_cameras
if num_cameras > 0
@@ -406,10 +431,10 @@ def record(
image_writer_processes=cfg.dataset.num_image_writer_processes,
image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera * len(robot.cameras),
batch_encoding_size=cfg.dataset.video_encoding_batch_size,
vcodec=cfg.dataset.vcodec,
camera_encoder=cfg.dataset.camera_encoder,
encoder_threads=cfg.dataset.encoder_threads,
streaming_encoding=cfg.dataset.streaming_encoding,
encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
encoder_threads=cfg.dataset.encoder_threads,
)
robot.connect()
@@ -420,7 +445,7 @@ def record(
if not cfg.dataset.streaming_encoding:
logging.info(
"Streaming encoding is disabled. If you have capable hardware, consider enabling it for way faster episode saving. --dataset.streaming_encoding=true --dataset.encoder_threads=2 # --dataset.vcodec=auto. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding"
"Streaming encoding is disabled. If you have capable hardware, consider enabling it for way faster episode saving. --dataset.streaming_encoding=true --dataset.encoder_threads=2 # --dataset.camera_encoder.vcodec=auto. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding"
)
with VideoEncodingManager(dataset):
src/lerobot/scripts/lerobot_replay.py
+2
View File
@@ -56,6 +56,7 @@ from lerobot.robots import ( # noqa: F401
Robot,
RobotConfig,
bi_openarm_follower,
bi_rebot_b601_follower,
bi_so_follower,
earthrover_mini_plus,
hope_jr,
@@ -64,6 +65,7 @@ from lerobot.robots import ( # noqa: F401
omx_follower,
openarm_follower,
reachy2,
rebot_b601_follower,
so_follower,
unitree_g1,
)
src/lerobot/scripts/lerobot_rollout.py
+16
View File
@@ -120,6 +120,18 @@ Usage examples
--dataset.repo_id=user/rollout_sentry_data \\
--dataset.single_task="patrol" \\
--resume=true
# Rollout with custom video encoding parameters
lerobot-rollout \\
--strategy.type=base \\
--policy.path=lerobot/act_koch_real \\
--robot.type=koch_follower \\
--robot.port=/dev/ttyACM0 \\
--task="pick up cube" --duration=60 \\
--display_data=true \\
--dataset.camera_encoder.vcodec=h264 \\
--dataset.camera_encoder.preset=fast \\
--dataset.camera_encoder.extra_options={"tune": "film", "profile:v": "high", "bf": 2}
"""
import logging
@@ -132,6 +144,7 @@ from lerobot.robots import ( # noqa: F401
Robot,
RobotConfig,
bi_openarm_follower,
bi_rebot_b601_follower,
bi_so_follower,
earthrover_mini_plus,
hope_jr,
@@ -139,6 +152,7 @@ from lerobot.robots import ( # noqa: F401
omx_follower,
openarm_follower,
reachy2,
rebot_b601_follower,
so_follower,
unitree_g1 as unitree_g1_robot,
)
@@ -147,6 +161,7 @@ from lerobot.teleoperators import ( # noqa: F401
Teleoperator,
TeleoperatorConfig,
bi_openarm_leader,
bi_rebot_102_leader,
bi_so_leader,
homunculus,
koch_leader,
@@ -154,6 +169,7 @@ from lerobot.teleoperators import ( # noqa: F401
openarm_leader,
openarm_mini,
reachy2_teleoperator,
rebot_102_leader,
so_leader,
unitree_g1,
)
src/lerobot/scripts/lerobot_setup_motors.py
+4
View File
@@ -30,20 +30,24 @@ import draccus
from lerobot.robots import ( # noqa: F401
RobotConfig,
bi_rebot_b601_follower,
bi_so_follower,
koch_follower,
lekiwi,
make_robot_from_config,
omx_follower,
rebot_b601_follower,
so_follower,
)
from lerobot.teleoperators import ( # noqa: F401
TeleoperatorConfig,
bi_rebot_102_leader,
bi_so_leader,
koch_leader,
make_teleoperator_from_config,
omx_leader,
openarm_mini,
rebot_102_leader,
so_leader,
)
src/lerobot/scripts/lerobot_teleoperate.py
+4
View File
@@ -72,6 +72,7 @@ from lerobot.robots import ( # noqa: F401
Robot,
RobotConfig,
bi_openarm_follower,
bi_rebot_b601_follower,
bi_so_follower,
earthrover_mini_plus,
hope_jr,
@@ -80,6 +81,7 @@ from lerobot.robots import ( # noqa: F401
omx_follower,
openarm_follower,
reachy2,
rebot_b601_follower,
so_follower,
unitree_g1 as unitree_g1_robot,
)
@@ -87,6 +89,7 @@ from lerobot.teleoperators import ( # noqa: F401
Teleoperator,
TeleoperatorConfig,
bi_openarm_leader,
bi_rebot_102_leader,
bi_so_leader,
gamepad,
homunculus,
@@ -97,6 +100,7 @@ from lerobot.teleoperators import ( # noqa: F401
openarm_leader,
openarm_mini,
reachy2_teleoperator,
rebot_102_leader,
so_leader,
unitree_g1,
)
src/lerobot/scripts/lerobot_train.py
+6
View File
@@ -48,6 +48,7 @@ from lerobot.envs import close_envs, make_env, make_env_pre_post_processors
from lerobot.optim.factory import make_optimizer_and_scheduler
from lerobot.policies import PreTrainedPolicy, make_policy, make_pre_post_processors
from lerobot.rewards import make_reward_pre_post_processors
from lerobot.utils.collate import lerobot_collate_fn
from lerobot.utils.import_utils import register_third_party_plugins
from lerobot.utils.logging_utils import AverageMeter, MetricsTracker
from lerobot.utils.random_utils import set_seed
@@ -401,6 +402,10 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
shuffle = True
sampler = None
# Only swap in the language-aware collate when the dataset actually
# declares language columns; otherwise stay on PyTorch's default
# collate so non-language training runs are unaffected.
collate_fn = lerobot_collate_fn if dataset.meta.has_language_columns else None
dataloader = torch.utils.data.DataLoader(
dataset,
num_workers=cfg.num_workers,
@@ -409,6 +414,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
sampler=sampler,
pin_memory=device.type == "cuda",
drop_last=False,
collate_fn=collate_fn,
prefetch_factor=cfg.prefetch_factor if cfg.num_workers > 0 else None,
persistent_workers=cfg.persistent_workers and cfg.num_workers > 0,
)
src/lerobot/teleoperators/bi_rebot_102_leader/__init__.py
+20
View File
@@ -0,0 +1,20 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .bi_rebot_102_leader import BiRebotArm102Leader
from .config_bi_rebot_102_leader import BiRebotArm102LeaderConfig
__all__ = ["BiRebotArm102Leader", "BiRebotArm102LeaderConfig"]
src/lerobot/teleoperators/bi_rebot_102_leader/bi_rebot_102_leader.py
+113
View File
@@ -0,0 +1,113 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
from functools import cached_property
from lerobot.types import RobotAction
from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
from ..rebot_102_leader import RebotArm102Leader, RebotArm102LeaderTeleopConfig
from ..teleoperator import Teleoperator
from .config_bi_rebot_102_leader import BiRebotArm102LeaderConfig
logger = logging.getLogger(__name__)
class BiRebotArm102Leader(Teleoperator):
"""Bimanual Seeed Studio StarArm102 / reBot Arm 102 leader.
Composes two single-arm :class:`RebotArm102Leader` instances. Action keys of
each arm are namespaced with a ``left_`` / ``right_`` prefix, so a bimanual
leader can teleoperate a bimanual reBot B601 follower.
"""
config_class = BiRebotArm102LeaderConfig
name = "bi_rebot_102_leader"
def __init__(self, config: BiRebotArm102LeaderConfig):
super().__init__(config)
self.config = config
left_arm_config = RebotArm102LeaderTeleopConfig(
id=f"{config.id}_left" if config.id else None,
calibration_dir=config.calibration_dir,
port=config.left_arm_config.port,
baudrate=config.left_arm_config.baudrate,
joint_ids=config.left_arm_config.joint_ids,
joint_directions=config.left_arm_config.joint_directions,
joint_ranges=config.left_arm_config.joint_ranges,
)
right_arm_config = RebotArm102LeaderTeleopConfig(
id=f"{config.id}_right" if config.id else None,
calibration_dir=config.calibration_dir,
port=config.right_arm_config.port,
baudrate=config.right_arm_config.baudrate,
joint_ids=config.right_arm_config.joint_ids,
joint_directions=config.right_arm_config.joint_directions,
joint_ranges=config.right_arm_config.joint_ranges,
)
self.left_arm = RebotArm102Leader(left_arm_config)
self.right_arm = RebotArm102Leader(right_arm_config)
@cached_property
def action_features(self) -> dict[str, type]:
return {
**{f"left_{k}": v for k, v in self.left_arm.action_features.items()},
**{f"right_{k}": v for k, v in self.right_arm.action_features.items()},
}
@cached_property
def feedback_features(self) -> dict[str, type]:
return {}
@property
def is_connected(self) -> bool:
return self.left_arm.is_connected and self.right_arm.is_connected
@check_if_already_connected
def connect(self, calibrate: bool = True) -> None:
self.left_arm.connect(calibrate)
self.right_arm.connect(calibrate)
@property
def is_calibrated(self) -> bool:
return self.left_arm.is_calibrated and self.right_arm.is_calibrated
def calibrate(self) -> None:
self.left_arm.calibrate()
self.right_arm.calibrate()
def configure(self) -> None:
self.left_arm.configure()
self.right_arm.configure()
@check_if_not_connected
def get_action(self) -> RobotAction:
action_dict = {}
action_dict.update({f"left_{k}": v for k, v in self.left_arm.get_action().items()})
action_dict.update({f"right_{k}": v for k, v in self.right_arm.get_action().items()})
return action_dict
def send_feedback(self, feedback: dict[str, float]) -> None:
raise NotImplementedError("Feedback is not implemented for the reBot Arm 102 leader.")
@check_if_not_connected
def disconnect(self) -> None:
self.left_arm.disconnect()
self.right_arm.disconnect()
src/lerobot/teleoperators/bi_rebot_102_leader/config_bi_rebot_102_leader.py
+29
View File
@@ -0,0 +1,29 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass
from ..config import TeleoperatorConfig
from ..rebot_102_leader import RebotArm102LeaderConfig
@TeleoperatorConfig.register_subclass("bi_rebot_102_leader")
@dataclass
class BiRebotArm102LeaderConfig(TeleoperatorConfig):
"""Configuration class for the bimanual reBot Arm 102 leader teleoperator."""
left_arm_config: RebotArm102LeaderConfig
right_arm_config: RebotArm102LeaderConfig
src/lerobot/teleoperators/rebot_102_leader/__init__.py
+20
View File
@@ -0,0 +1,20 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .config_rebot_102_leader import RebotArm102LeaderConfig, RebotArm102LeaderTeleopConfig
from .rebot_102_leader import RebotArm102Leader
__all__ = ["RebotArm102Leader", "RebotArm102LeaderConfig", "RebotArm102LeaderTeleopConfig"]
src/lerobot/teleoperators/rebot_102_leader/config_rebot_102_leader.py
+83
View File
@@ -0,0 +1,83 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass, field
from ..config import TeleoperatorConfig
@dataclass
class RebotArm102LeaderConfig:
"""Base configuration class for the Seeed Studio StarArm102 / reBot Arm 102 leader.
The reBot Arm 102 is a 7-joint (incl. gripper) leader arm driven by FashionStar
UART smart servos. Servo communication goes through ``motorbridge-smart-servo``.
"""
# USB-to-UART device the leader arm is connected to (e.g. "/dev/ttyUSB0").
port: str
baudrate: int = 1_000_000
# Servo id of each joint on the UART bus.
joint_ids: dict[str, int] = field(
default_factory=lambda: {
"shoulder_pan": 0,
"shoulder_lift": 1,
"elbow_flex": 2,
"wrist_flex": 3,
"wrist_yaw": 4,
"wrist_roll": 5,
"gripper": 6,
}
)
# Per-joint sign applied to raw servo angles so the leader matches the follower
# convention. The gripper additionally carries a scale (e.g. -6) to widen its
# range to the reBot B601 follower's gripper travel.
joint_directions: dict[str, int] = field(
default_factory=lambda: {
"shoulder_pan": -1,
"shoulder_lift": -1,
"elbow_flex": 1,
"wrist_flex": 1,
"wrist_yaw": 1,
"wrist_roll": -1,
"gripper": -6,
}
)
# Per-joint [min, max] output range in degrees. Matches the reBot B601 follower
# joint limits so leader actions can drive the follower key-for-key.
joint_ranges: dict[str, list[int]] = field(
default_factory=lambda: {
"shoulder_pan": [-150, 150],
"shoulder_lift": [-170, 1],
"elbow_flex": [-200, 1],
"wrist_flex": [-80, 90],
"wrist_yaw": [-90, 90],
"wrist_roll": [-90, 90],
"gripper": [-270, 0],
}
)
@TeleoperatorConfig.register_subclass("rebot_102_leader")
@dataclass
class RebotArm102LeaderTeleopConfig(TeleoperatorConfig, RebotArm102LeaderConfig):
"""Registered configuration for the reBot Arm 102 leader teleoperator."""
pass
src/lerobot/teleoperators/rebot_102_leader/rebot_102_leader.py
+207
View File
@@ -0,0 +1,207 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
import time
from typing import TYPE_CHECKING
from lerobot.motors import MotorCalibration
from lerobot.types import RobotAction
from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
from lerobot.utils.import_utils import _motorbridge_smart_servo_available, require_package
from ..teleoperator import Teleoperator
from .config_rebot_102_leader import RebotArm102LeaderTeleopConfig
if TYPE_CHECKING or _motorbridge_smart_servo_available:
from motorbridge_smart_servo import FashionStarServo, ServoMonitor
else:
FashionStarServo = None
ServoMonitor = None
logger = logging.getLogger(__name__)
_SETTLE_SEC = 0.01
class RebotArm102Leader(Teleoperator):
"""Seeed Studio StarArm102 / reBot Arm 102 leader arm.
A 7-joint (incl. gripper) leader built on FashionStar UART smart servos. Servo
communication is handled by the ``motorbridge-smart-servo`` package; this class
only reads joint angles, so it produces actions but accepts no feedback.
"""
config_class = RebotArm102LeaderTeleopConfig
name = "rebot_102_leader"
def __init__(self, config: RebotArm102LeaderTeleopConfig):
require_package("motorbridge-smart-servo", extra="rebot", import_name="motorbridge_smart_servo")
super().__init__(config)
self.config = config
self.bus: FashionStarServo | None = None
self.motor_names = list(config.joint_ids.keys())
self._last_raw_positions: dict[str, float] = {}
@property
def action_features(self) -> dict[str, type]:
return {f"{motor}.pos": float for motor in self.motor_names}
@property
def feedback_features(self) -> dict[str, type]:
return {}
@property
def is_connected(self) -> bool:
return self.bus is not None
@check_if_already_connected
def connect(self, calibrate: bool = True) -> None:
logger.info(f"Connecting {self} on {self.config.port}...")
bus = FashionStarServo(self.config.port, baudrate=self.config.baudrate)
try:
for motor_name, motor_id in self.config.joint_ids.items():
if not bus.ping(motor_id):
raise RuntimeError(f"Servo not found for {motor_name} (id={motor_id}).")
self._last_raw_positions[motor_name] = 0.0
self.bus = bus
if not self.is_calibrated and calibrate:
logger.info(
"Mismatch between calibration values in the motor and the calibration file or no calibration file found"
)
self.calibrate()
self.configure()
except Exception:
bus.close()
self.bus = None
raise
logger.info(f"{self} connected.")
@property
def is_calibrated(self) -> bool:
return bool(self.calibration) and set(self.calibration) == set(self.motor_names)
def calibrate(self) -> None:
if self.calibration:
user_input = input(
f"Press ENTER to use provided calibration file associated with the id {self.id}, "
"or type 'c' and press ENTER to run calibration: "
)
if user_input.strip().lower() != "c":
logger.info(f"Using calibration file associated with the id {self.id}")
return
logger.info(f"\nRunning calibration of {self}")
input(
"\nCalibration: set zero position.\n"
"Manually move the reBot Arm 102 to its zero pose and close the gripper.\n"
"Press ENTER when ready..."
)
self.calibration = {}
for motor_name, motor_id in self.config.joint_ids.items():
self.bus.unlock(motor_id)
time.sleep(_SETTLE_SEC)
self.bus.set_origin_point(motor_id)
range_min, range_max = self.config.joint_ranges[motor_name]
self.calibration[motor_name] = MotorCalibration(
id=motor_id,
drive_mode=0,
homing_offset=0,
range_min=int(range_min),
range_max=int(range_max),
)
self._save_calibration()
logger.info(f"Calibration saved to {self.calibration_fpath}")
def configure(self) -> None:
for motor_id in self.config.joint_ids.values():
self.bus.unlock(motor_id)
time.sleep(_SETTLE_SEC)
# Reset the multi-turn counter of each servo individually.
for motor_id in self.config.joint_ids.values():
self.bus.reset_multi_turn(motor_id)
def _read_raw_positions(self) -> dict[str, float]:
result: dict[int, ServoMonitor | None] = self.bus.sync_monitor(list(self.config.joint_ids.values()))
id_to_name = {v: k for k, v in self.config.joint_ids.items()}
raw_positions: dict[str, float] = {}
for motor_id, monitor in result.items():
motor_name = id_to_name[motor_id]
if monitor is None:
raise RuntimeError(f"Servo {motor_name} (id={motor_id}) has never responded.")
raw_positions[motor_name] = monitor.angle_deg
return raw_positions
@staticmethod
def _round_to_valid_range(value: float, min_value: float, max_value: float) -> tuple[float, int]:
"""Unwrap a multi-turn angle into the ±180° window centred on (min+max)/2.
The servo may report an angle that has accumulated extra full rotations
(value = true_angle + N*360). Subtract the nearest whole number of turns
to bring it back into [center-180, center+180]. Returns the unwrapped
angle and the number of turns removed.
"""
center = (min_value + max_value) / 2.0
turns = round((value - center) / 360.0)
return value - turns * 360.0, abs(turns)
@check_if_not_connected
def get_action(self) -> RobotAction:
start = time.perf_counter()
try:
raw_positions = self._read_raw_positions()
self._last_raw_positions = raw_positions
except Exception as e:
logger.error(f"Failed to read raw positions: {e}")
logger.warning("[EMERGENCY STOP] Hold the follower arm and cut off the main power to the arms.")
logger.warning(
"[EMERGENCY STOP] Break the teleoperation session and check the leader USB connection or power."
)
raw_positions = self._last_raw_positions
action_dict: dict[str, float] = {}
for motor_name in self.motor_names:
range_min, range_max = self.config.joint_ranges[motor_name]
direction = self.config.joint_directions[motor_name]
sign = 1.0 if direction >= 0 else -1.0
unwrapped, k = self._round_to_valid_range(
raw_positions[motor_name], range_min * sign, range_max * sign
)
position = unwrapped * direction
if k > 0:
logger.debug(
f"Servo {motor_name} (id={self.config.joint_ids[motor_name]}) wrapped {k} * 360°. "
f"Unwrapped pos: {unwrapped:.1f}° (raw: {raw_positions[motor_name]:.1f}°)"
)
action_dict[f"{motor_name}.pos"] = max(float(range_min), min(float(range_max), position))
dt_ms = (time.perf_counter() - start) * 1e3
logger.debug(f"{self} read action: {dt_ms:.1f}ms")
return action_dict
def send_feedback(self, feedback: dict[str, float]) -> None:
raise NotImplementedError("Feedback is not implemented for the reBot Arm 102 leader.")
@check_if_not_connected
def disconnect(self) -> None:
self.bus.close()
self.bus = None
logger.info(f"{self} disconnected.")
src/lerobot/teleoperators/utils.py
+8
View File
@@ -99,6 +99,14 @@ def make_teleoperator_from_config(config: TeleoperatorConfig) -> "Teleoperator":
from .openarm_mini import OpenArmMini
return OpenArmMini(config)
elif config.type == "rebot_102_leader":
from .rebot_102_leader import RebotArm102Leader
return RebotArm102Leader(config)
elif config.type == "bi_rebot_102_leader":
from .bi_rebot_102_leader import BiRebotArm102Leader
return BiRebotArm102Leader(config)
else:
try:
return cast("Teleoperator", make_device_from_device_class(config))
src/lerobot/templates/lerobot_modelcard_template.md
-2
View File
@@ -41,8 +41,6 @@ For more details, see the [Physical Intelligence π₀ blog post](https://www.ph
For more details, see the [Physical Intelligence π₀.₅ blog post](https://www.physicalintelligence.company/blog/pi05).
{% elif model_name == "gaussian_actor" %}
This is a Gaussian Actor policy (Gaussian policy with a tanh squash) — the policy-side component used by [Soft Actor-Critic (SAC)](https://huggingface.co/papers/1801.01290) and related maximum-entropy continuous-control algorithms.
{% elif model_name == "reward_classifier" %}
A reward classifier is a lightweight neural network that scores observations or trajectories for task success, providing a learned reward signal or offline evaluation when explicit rewards are unavailable.
{% else %}
_Model type not recognized — please update this template._
{% endif %}
src/lerobot/transport/utils.py
+2 -1
View File
@@ -25,9 +25,10 @@ from typing import Any
import torch
from lerobot.transport import services_pb2
from lerobot.utils.transition import Transition
from . import services_pb2
# FIX for protobuf: Assign the enum to a variable and ignore the type error once
TransferState = services_pb2.TransferState # type: ignore[attr-defined]
src/lerobot/utils/collate.py
+65
View File
@@ -0,0 +1,65 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations
from typing import Any
from torch.utils.data._utils.collate import default_collate
from lerobot.datasets.language import LANGUAGE_COLUMNS
_PYTHON_LIST_KEYS = {"messages", "message_streams", "target_message_indices"}
def lerobot_collate_fn(batch: list[dict[str, Any] | None]) -> dict[str, Any] | None:
"""Collate function that preserves Python-list and language fields as lists.
Drops ``None`` samples (e.g. recipes that yielded no target message), keeps
rendered-message and language fields as plain Python lists, and delegates
every other key to PyTorch's ``default_collate``.
"""
batch = [sample for sample in batch if sample is not None]
if not batch:
return None
# All-or-nothing per key: a partial-presence batch (e.g. half the samples
# carry `messages` and half don't) is a real bug in the upstream
# rendering step — silently filtering would hand downstream consumers a
# preserved list shorter than the tensor batch. Raise instead so the
# mismatch surfaces at the boundary.
preserved: dict[str, list[Any]] = {}
for key in _PYTHON_LIST_KEYS:
presence = [key in sample for sample in batch]
if not any(presence):
continue
if not all(presence):
raise ValueError(
f"Inconsistent batch: {sum(presence)}/{len(batch)} samples carry {key!r}; "
f"every sample in a batch must agree."
)
preserved[key] = [sample[key] for sample in batch]
tensorizable = [
{
key: value
for key, value in sample.items()
if key not in _PYTHON_LIST_KEYS and key not in LANGUAGE_COLUMNS
}
for sample in batch
]
collated = default_collate(tensorizable)
collated.update(preserved)
return collated
src/lerobot/utils/import_utils.py
+8 -1
View File
@@ -69,7 +69,7 @@ def is_package_available(
return package_exists
def get_safe_default_codec():
def get_safe_default_video_backend():
logger = logging.getLogger(__name__)
if importlib.util.find_spec("torchcodec"):
return "torchcodec"
@@ -114,6 +114,10 @@ _dynamixel_sdk_available = is_package_available("dynamixel-sdk", import_name="dy
_feetech_sdk_available = is_package_available("feetech-servo-sdk", import_name="scservo_sdk")
_reachy2_sdk_available = is_package_available("reachy2_sdk")
_can_available = is_package_available("python-can", "can")
_motorbridge_available = is_package_available("motorbridge")
_motorbridge_smart_servo_available = is_package_available(
"motorbridge-smart-servo", import_name="motorbridge_smart_servo"
)
_unitree_sdk_available = is_package_available("unitree-sdk2py", "unitree_sdk2py")
_pyrealsense2_available = is_package_available("pyrealsense2") or is_package_available(
"pyrealsense2-macosx", import_name="pyrealsense2"
@@ -128,6 +132,9 @@ _hidapi_available = is_package_available("hidapi", import_name="hid")
_pandas_available = is_package_available("pandas")
_faker_available = is_package_available("faker")
# Video encoding / decoding
_av_available = is_package_available("av")
# Misc
_pynput_available = is_package_available("pynput")
_pygame_available = is_package_available("pygame")
src/lerobot/utils/utils.py
+19
View File
@@ -160,6 +160,25 @@ def has_method(cls: object, method_name: str) -> bool:
return hasattr(cls, method_name) and callable(getattr(cls, method_name))
def unwrap_scalar(value: Any) -> Any:
"""Unwrap a tensor / numpy scalar / single-element list into a Python scalar.
Tensors and numpy scalars expose ``.item()``; single-element lists are
unwrapped recursively. Anything else is returned unchanged. Centralized
here so the language renderer and processor steps share one definition.
Raises:
ValueError: If ``value`` is a list with zero or multiple elements.
"""
if hasattr(value, "item"):
return value.item()
if isinstance(value, list):
if len(value) != 1:
raise ValueError(f"Expected a scalar, got list of length {len(value)}: {value!r}")
return unwrap_scalar(value[0])
return value
def is_valid_numpy_dtype_string(dtype_str: str) -> bool:
"""
Return True if a given string can be converted to a numpy dtype.
tests/artifacts/encoded_videos/clip_32x48.mp4
+3
View File
@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:2191cd86e9e32ecbe18e33ad68d49060e479723ab5a3212bbb26df3025ccb568
size 5815
tests/artifacts/encoded_videos/clip_4frames.mp4
+3
View File
@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3e0ebf563ba3ed9c24b691a0f0b29e0294a1fa9b51422e1ece296155f1465768
size 16236
tests/artifacts/encoded_videos/clip_5frames.mp4
+3
View File
@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8475bfd5e6c4c780df46200e2b027e262b38436c57d01078bd943a5b87c65b8f
size 20726
tests/artifacts/encoded_videos/clip_6frames.mp4
+3
View File
@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:6434322d1c671a7d132367619f841a775317cb9ff973f3f4505831e3ed74076d
size 23808
tests/artifacts/encoded_videos/clip_h264.mp4
+3
View File
@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8efc84375e92a3499cef93100e04d8fb354670f3d9e0db2097b52575927284fc
size 12237
tests/configs/test_recipe.py
+168
View File
@@ -0,0 +1,168 @@
#!/usr/bin/env python
from pathlib import Path
from textwrap import dedent
import pytest
from lerobot.configs.recipe import MessageTurn, TrainingRecipe, load_recipe
def _minimal_message_turn(content: str = "${task}") -> MessageTurn:
return MessageTurn(role="user", content=content, stream="high_level")
def _minimal_target_turn() -> MessageTurn:
return MessageTurn(role="assistant", content="ok", stream="high_level", target=True)
# ── Message-recipe validation ────────────────────────────────────────
def test_message_recipe_validates_unknown_binding():
with pytest.raises(ValueError, match="unknown binding"):
TrainingRecipe(
messages=[
MessageTurn(role="user", content="${missing}", stream="high_level"),
_minimal_target_turn(),
]
)
def test_message_turn_requires_a_stream():
"""Every turn must declare a stream — None is rejected at construction.
Previously this only failed at render time (``_validate_rendered``);
catching it here means a malformed recipe YAML errors at load instead
of at the first training sample.
"""
with pytest.raises(ValueError, match="missing a stream"):
MessageTurn(role="user", content="${task}")
def test_message_recipe_requires_at_least_one_target():
with pytest.raises(ValueError, match="target"):
TrainingRecipe(
messages=[
_minimal_message_turn(),
MessageTurn(role="assistant", content="no target", stream="high_level"),
]
)
def test_recipe_rejects_both_messages_and_blend():
with pytest.raises(ValueError, match="only one"):
TrainingRecipe(
messages=[_minimal_message_turn(), _minimal_target_turn()],
blend={"a": TrainingRecipe(weight=1.0, messages=[_minimal_target_turn()])},
)
def test_recipe_rejects_neither_messages_nor_blend():
with pytest.raises(ValueError, match="must set one"):
TrainingRecipe()
# ── Blend validation ─────────────────────────────────────────────────
def test_blend_must_be_non_empty():
with pytest.raises(ValueError, match="at least one component"):
TrainingRecipe(blend={})
def test_blend_component_must_define_weight():
with pytest.raises(ValueError, match="weight"):
TrainingRecipe(blend={"a": TrainingRecipe(messages=[_minimal_target_turn()])})
def test_blend_component_weight_must_be_positive():
with pytest.raises(ValueError, match="positive weight"):
TrainingRecipe(blend={"a": TrainingRecipe(weight=0.0, messages=[_minimal_target_turn()])})
def test_blend_component_must_define_messages():
# A bare TrainingRecipe(weight=1.0) would itself raise; build it without
# going through __post_init__ to exercise the blend-level validator.
bad = TrainingRecipe.__new__(TrainingRecipe)
bad.messages = None
bad.bindings = None
bad.blend = None
bad.weight = 1.0
with pytest.raises(ValueError, match="must define messages"):
TrainingRecipe(blend={"a": bad})
def test_blend_components_cannot_themselves_define_a_blend():
inner = TrainingRecipe(blend={"x": TrainingRecipe(weight=1.0, messages=[_minimal_target_turn()])})
# Force-bypass the inner component's normal validation so the test
# exercises the outer blend's "no nested blends" rule directly.
nested = TrainingRecipe.__new__(TrainingRecipe)
nested.messages = None
nested.bindings = None
nested.blend = inner.blend
nested.weight = 1.0
with pytest.raises(ValueError, match="cannot itself define a blend"):
TrainingRecipe(blend={"outer": nested})
# ── from_dict / from_yaml round-trips ────────────────────────────────
def test_from_dict_with_nested_blend():
recipe = TrainingRecipe.from_dict(
{
"blend": {
"a": {
"weight": 1.0,
"messages": [
{"role": "user", "content": "${task}", "stream": "high_level"},
{"role": "assistant", "content": "a", "stream": "high_level", "target": True},
],
},
"b": {
"weight": 2.0,
"messages": [
{"role": "user", "content": "${task}", "stream": "high_level"},
{"role": "assistant", "content": "b", "stream": "high_level", "target": True},
],
},
}
}
)
assert recipe.blend is not None
assert set(recipe.blend) == {"a", "b"}
assert recipe.blend["b"].weight == 2.0
# Inner messages were promoted to MessageTurn instances.
assert isinstance(recipe.blend["a"].messages[0], MessageTurn)
def test_from_yaml_round_trips_through_load_recipe(tmp_path: Path):
yaml_text = dedent(
"""
bindings:
custom: "active_at(t, style=subtask)"
messages:
- {role: user, content: "${task}: ${custom}", stream: high_level}
- {role: assistant, content: "ok", stream: high_level, target: true}
"""
).strip()
path = tmp_path / "recipe.yaml"
path.write_text(yaml_text)
via_classmethod = TrainingRecipe.from_yaml(path)
via_helper = load_recipe(path)
assert via_classmethod.bindings == {"custom": "active_at(t, style=subtask)"}
assert via_classmethod.messages[1].target is True
# ``load_recipe`` is just a wrapper, but assert the two paths agree
# on the structural result so a future divergence is caught here.
assert via_helper.bindings == via_classmethod.bindings
assert len(via_helper.messages) == len(via_classmethod.messages)
def test_from_yaml_rejects_non_mapping(tmp_path: Path):
path = tmp_path / "bad.yaml"
path.write_text("- just\n- a\n- list\n")
with pytest.raises(ValueError, match="mapping at the top level"):
TrainingRecipe.from_yaml(path)
tests/datasets/test_aggregate.py
+74 -2
View File
@@ -14,6 +14,8 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import json
import logging
from unittest.mock import patch
import pytest
@@ -23,7 +25,9 @@ pytest.importorskip("datasets", reason="datasets is required (install lerobot[da
import datasets # noqa: E402
import torch
from lerobot.configs import VIDEO_ENCODER_INFO_KEYS
from lerobot.datasets.aggregate import aggregate_datasets
from lerobot.datasets.feature_utils import features_equal_for_merge
from lerobot.datasets.lerobot_dataset import LeRobotDataset
from tests.fixtures.constants import DUMMY_REPO_ID
@@ -117,8 +121,9 @@ def assert_metadata_consistency(aggr_ds, ds_0, ds_1):
"Robot type should be the same"
)
# Test features are the same
assert aggr_ds.features == ds_0.features == ds_1.features, "Features should be the same"
# Schema matches; merged video ``info`` is reconciled separately from per-source ``info``.
assert features_equal_for_merge(aggr_ds.features, ds_0.features)
assert features_equal_for_merge(aggr_ds.features, ds_1.features)
# Test tasks aggregation
expected_tasks = set(ds_0.meta.tasks.index) | set(ds_1.meta.tasks.index)
@@ -284,6 +289,73 @@ def test_aggregate_datasets(tmp_path, lerobot_dataset_factory):
assert_dataset_iteration_works(aggr_ds)
@pytest.mark.parametrize("mutation", ["mismatched_value", "missing_key"])
def test_aggregate_incomplete_video_encoder_info_warns_and_nuls_encoders(
tmp_path, lerobot_dataset_factory, caplog, mutation
):
"""Mismatched or missing encoder ``info`` is merged per-key with fallbacks and a warning."""
suffix = "enc_mismatch" if mutation == "mismatched_value" else "enc_missing"
ds_0 = lerobot_dataset_factory(
root=tmp_path / f"{suffix}_a",
repo_id=f"{DUMMY_REPO_ID}_{suffix}_a",
total_episodes=2,
total_frames=20,
)
ds_1 = lerobot_dataset_factory(
root=tmp_path / f"{suffix}_b",
repo_id=f"{DUMMY_REPO_ID}_{suffix}_b",
total_episodes=2,
total_frames=20,
)
info_path = ds_1.root / "meta" / "info.json"
data = json.loads(info_path.read_text())
for ft in data["features"].values():
if ft.get("dtype") != "video":
continue
inf = ft.setdefault("info", {})
if mutation == "mismatched_value":
inf["video.crf"] = 99
inf["video.extra_options"] = {"tune": "film"}
else:
inf.pop("video.crf", None)
inf.pop("video.extra_options", None)
info_path.write_text(json.dumps(data))
aggr_id = f"{DUMMY_REPO_ID}_{suffix}_aggr"
aggr_root = tmp_path / f"{suffix}_aggr"
with caplog.at_level(logging.WARNING):
aggregate_datasets(
repo_ids=[ds_0.repo_id, ds_1.repo_id],
roots=[ds_0.root, ds_1.root],
aggr_repo_id=aggr_id,
aggr_root=aggr_root,
)
assert "heterogeneous" in caplog.text.lower() or "incomplete" in caplog.text.lower()
with (
patch("lerobot.datasets.dataset_metadata.get_safe_version") as mock_get_safe_version,
patch("lerobot.datasets.dataset_metadata.snapshot_download") as mock_snapshot_download,
):
mock_get_safe_version.return_value = "v3.0"
mock_snapshot_download.return_value = str(aggr_root)
aggr_ds = LeRobotDataset(aggr_id, root=aggr_root)
for key, ft in aggr_ds.meta.info.features.items():
if ft.get("dtype") != "video":
continue
info = ft["info"]
reference = ds_0.meta.info.features[key]["info"]
for info_key in VIDEO_ENCODER_INFO_KEYS:
if info_key == "video.crf":
assert info[info_key] is None
elif info_key == "video.extra_options":
assert info[info_key] == {}
else:
assert info[info_key] == reference[info_key]
def test_aggregate_with_low_threshold(tmp_path, lerobot_dataset_factory):
"""Test aggregation with small file size limits to force file rotation/sharding."""
ds_0_num_episodes = ds_1_num_episodes = 10
tests/datasets/test_dataset_metadata.py
+137
View File
@@ -385,3 +385,140 @@ def test_finalize_flushes_buffered_metadata(tmp_path):
assert episodes_dir.exists()
parquet_files = list(episodes_dir.rglob("*.parquet"))
assert len(parquet_files) > 0
# ── Tools accessor ───────────────────────────────────────────────────
def test_tools_falls_back_to_default_when_info_has_no_tools_field(tmp_path):
"""meta.tools returns DEFAULT_TOOLS when info.json doesn't declare any."""
from lerobot.datasets.language import DEFAULT_TOOLS
root = tmp_path / "no_tools"
meta = LeRobotDatasetMetadata.create(
repo_id="test/no_tools",
fps=DEFAULT_FPS,
features=SIMPLE_FEATURES,
root=root,
use_videos=False,
)
assert meta.tools == DEFAULT_TOOLS
# info.json on disk should NOT include a `tools` key for clean datasets
with open(root / INFO_PATH) as f:
info_on_disk = json.load(f)
assert "tools" not in info_on_disk
def test_tools_reads_declared_tools_from_info_json(tmp_path):
"""A `tools` list written into info.json survives load → meta.tools.
Regression test for the bug where ``DatasetInfo.from_dict`` silently
dropped the ``tools`` key (no matching dataclass field), so
``meta.tools`` always returned ``DEFAULT_TOOLS`` regardless of
what was on disk.
"""
from lerobot.datasets.io_utils import load_info
root = tmp_path / "with_tools"
meta = LeRobotDatasetMetadata.create(
repo_id="test/with_tools",
fps=DEFAULT_FPS,
features=SIMPLE_FEATURES,
root=root,
use_videos=False,
)
custom_tool = {
"type": "function",
"function": {
"name": "record_observation",
"description": "Capture a still image.",
"parameters": {
"type": "object",
"properties": {"label": {"type": "string"}},
"required": ["label"],
},
},
}
info_path = root / INFO_PATH
with open(info_path) as f:
raw = json.load(f)
raw["tools"] = [custom_tool]
with open(info_path, "w") as f:
json.dump(raw, f)
# Reload info from disk and rebind it on the metadata object
meta.info = load_info(root)
assert meta.tools == [custom_tool]
def test_tools_round_trip_through_dataset_info(tmp_path):
"""A `tools` list survives DatasetInfo.from_dict / to_dict."""
from lerobot.datasets.utils import DatasetInfo
raw = {
"codebase_version": "v3.1",
"fps": 30,
"features": SIMPLE_FEATURES,
"tools": [{"type": "function", "function": {"name": "say"}}],
}
info = DatasetInfo.from_dict(raw)
assert info.tools == raw["tools"]
assert info.to_dict()["tools"] == raw["tools"]
def test_tools_setter_persists_to_info_json_and_reloads(tmp_path):
"""Assigning meta.tools writes info.json and reloads meta.info."""
from lerobot.datasets.io_utils import load_info
root = tmp_path / "set_tools"
meta = LeRobotDatasetMetadata.create(
repo_id="test/set_tools",
fps=DEFAULT_FPS,
features=SIMPLE_FEATURES,
root=root,
use_videos=False,
)
custom_tool = {
"type": "function",
"function": {
"name": "record_observation",
"description": "Capture a still image.",
"parameters": {
"type": "object",
"properties": {"label": {"type": "string"}},
"required": ["label"],
},
},
}
meta.tools = [custom_tool]
# In-memory metadata reflects the new catalog ...
assert meta.tools == [custom_tool]
assert meta.info.tools == [custom_tool]
# ... and a fresh read from disk agrees.
assert load_info(root).tools == [custom_tool]
def test_tools_setter_clears_key_when_set_to_none(tmp_path):
"""Setting meta.tools back to None drops the key and restores the default."""
from lerobot.datasets.language import DEFAULT_TOOLS
root = tmp_path / "clear_tools"
meta = LeRobotDatasetMetadata.create(
repo_id="test/clear_tools",
fps=DEFAULT_FPS,
features=SIMPLE_FEATURES,
root=root,
use_videos=False,
)
meta.tools = [{"type": "function", "function": {"name": "say"}}]
meta.tools = None
assert meta.tools == DEFAULT_TOOLS
with open(root / INFO_PATH) as f:
info_on_disk = json.load(f)
assert "tools" not in info_on_disk
tests/datasets/test_dataset_reader.py
+3 -3
View File
@@ -20,7 +20,7 @@ import pytest
pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
from lerobot.datasets.dataset_reader import DatasetReader
from lerobot.utils.import_utils import get_safe_default_codec
from lerobot.utils.import_utils import get_safe_default_video_backend
# ── Loading ──────────────────────────────────────────────────────────
@@ -35,7 +35,7 @@ def test_try_load_returns_true_when_data_exists(tmp_path, lerobot_dataset_factor
root=dataset.root,
episodes=None,
tolerance_s=1e-4,
video_backend=get_safe_default_codec(),
video_backend=get_safe_default_video_backend(),
delta_timestamps=None,
image_transforms=None,
)
@@ -58,7 +58,7 @@ def test_try_load_returns_false_when_no_data(tmp_path):
root=meta.root,
episodes=None,
tolerance_s=1e-4,
video_backend=get_safe_default_codec(),
video_backend=get_safe_default_video_backend(),
delta_timestamps=None,
image_transforms=None,
)
tests/datasets/test_dataset_tools.py
+50 -5
View File
@@ -23,16 +23,21 @@ import torch
pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
from lerobot.configs import VideoEncoderConfig
from lerobot.datasets.dataset_tools import (
add_features,
convert_image_to_video_dataset,
delete_episodes,
merge_datasets,
modify_features,
modify_tasks,
reencode_dataset,
remove_feature,
split_dataset,
)
from lerobot.scripts.lerobot_edit_dataset import convert_image_to_video_dataset
from lerobot.datasets.io_utils import load_info
from tests.datasets.test_video_encoding import _add_frames, require_h264, require_libsvtav1
@pytest.fixture
@@ -1246,10 +1251,12 @@ def test_convert_image_to_video_dataset(tmp_path):
dataset=source_dataset,
output_dir=output_dir,
repo_id="lerobot/pusht_video",
vcodec="libsvtav1",
pix_fmt="yuv420p",
g=2,
crf=30,
camera_encoder=VideoEncoderConfig(
vcodec="libsvtav1",
pix_fmt="yuv420p",
g=2,
crf=30,
),
episode_indices=[0, 1],
num_workers=2,
)
@@ -1323,3 +1330,41 @@ def test_convert_image_to_video_dataset_subset_episodes(tmp_path):
if output_dir.exists():
shutil.rmtree(output_dir)
# ─── reencode_dataset ─────────────────────────────────────────────────
@require_libsvtav1
@require_h264
def test_reencode_dataset_multi_key_multiprocessing(
tmp_path, empty_lerobot_dataset_factory, features_factory
):
"""Re-encode a two-camera dataset with num_workers=2 and verify metadata refresh."""
features = features_factory(use_videos=True)
initial_cfg = VideoEncoderConfig(vcodec="libsvtav1", g=2, crf=30, preset=12)
dataset = empty_lerobot_dataset_factory(
root=tmp_path / "ds",
features=features,
use_videos=True,
camera_encoder=initial_cfg,
)
_add_frames(dataset, num_frames=4)
dataset.save_episode()
_add_frames(dataset, num_frames=4)
dataset.save_episode()
dataset.finalize()
assert len(dataset.meta.video_keys) == 2
target_cfg = VideoEncoderConfig(vcodec="h264", g=6, crf=23, pix_fmt="yuv420p")
result = reencode_dataset(dataset, camera_encoder=target_cfg, num_workers=2)
assert result is dataset
persisted_info = load_info(dataset.root)
for vk in dataset.meta.video_keys:
persisted_encoder = VideoEncoderConfig.from_video_info(persisted_info.features[vk].get("info", {}))
assert persisted_encoder == target_cfg
tests/datasets/test_dataset_writer.py
+17 -7
View File
@@ -25,6 +25,7 @@ from PIL import Image
pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
from lerobot.configs import VideoEncoderConfig
from lerobot.datasets.dataset_writer import _encode_video_worker
from lerobot.datasets.lerobot_dataset import LeRobotDataset
from lerobot.datasets.utils import DEFAULT_IMAGE_PATH
@@ -52,8 +53,8 @@ def _make_frame(features: dict, task: str = "Dummy task") -> dict:
# ── Existing encode_video_worker tests ───────────────────────────────
def test_encode_video_worker_forwards_vcodec(tmp_path):
"""_encode_video_worker correctly forwards the vcodec parameter."""
def test_encode_video_worker_forwards_camera_encoder(tmp_path):
"""_encode_video_worker forwards camera_encoder to encode_video_frames."""
video_key = "observation.images.laptop"
fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=0, frame_index=0)
img_dir = tmp_path / Path(fpath).parent
@@ -68,13 +69,21 @@ def test_encode_video_worker_forwards_vcodec(tmp_path):
Path(video_path).touch()
with patch("lerobot.datasets.dataset_writer.encode_video_frames", side_effect=mock_encode):
_encode_video_worker(video_key, 0, tmp_path, fps=30, vcodec="h264")
_encode_video_worker(
video_key,
0,
tmp_path,
fps=30,
camera_encoder=VideoEncoderConfig(vcodec="h264", preset=None),
encoder_threads=4,
)
assert captured_kwargs["vcodec"] == "h264"
assert captured_kwargs["camera_encoder"].vcodec == "h264"
assert captured_kwargs["encoder_threads"] == 4
def test_encode_video_worker_default_vcodec(tmp_path):
"""_encode_video_worker uses libsvtav1 as the default codec."""
def test_encode_video_worker_default_camera_encoder(tmp_path):
"""_encode_video_worker passes None camera_encoder which encode_video_frames defaults."""
video_key = "observation.images.laptop"
fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=0, frame_index=0)
img_dir = tmp_path / Path(fpath).parent
@@ -91,7 +100,8 @@ def test_encode_video_worker_default_vcodec(tmp_path):
with patch("lerobot.datasets.dataset_writer.encode_video_frames", side_effect=mock_encode):
_encode_video_worker(video_key, 0, tmp_path, fps=30)
assert captured_kwargs["vcodec"] == "libsvtav1"
assert captured_kwargs["camera_encoder"] is None
assert captured_kwargs["encoder_threads"] is None
# ── add_frame contracts ──────────────────────────────────────────────
tests/datasets/test_datasets.py
+39 -11
View File
@@ -24,11 +24,13 @@ import torch
pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
import datasets
from huggingface_hub import HfApi
from PIL import Image
from safetensors.torch import load_file
from torchvision.transforms import v2
from lerobot.configs import VALID_VIDEO_CODECS, VideoEncoderConfig
from lerobot.configs.default import DatasetConfig
from lerobot.configs.train import TrainPipelineConfig
from lerobot.datasets import make_dataset
@@ -43,7 +45,6 @@ from lerobot.datasets.utils import (
DEFAULT_VIDEO_FILE_SIZE_IN_MB,
create_branch,
)
from lerobot.datasets.video_utils import VALID_VIDEO_CODECS
from lerobot.envs.factory import make_env_config
from lerobot.policies.factory import make_policy_config
from lerobot.robots import make_robot_from_config
@@ -360,6 +361,41 @@ def test_add_frame_image_pil(image_dataset):
assert dataset[0]["image"].shape == torch.Size(DUMMY_CHW)
@pytest.mark.parametrize(
"dtype,np_dtype,values,assert_fn",
[
("float32", np.float32, [1.0, 2.0], np.testing.assert_allclose),
("int64", np.int64, [1, 2], np.testing.assert_array_equal),
("bool", np.bool_, [True, False], np.testing.assert_array_equal),
],
ids=["float32", "int64", "bool"],
)
def test_save_episode_shape_1_scalar_is_scalarized_before_hf_encoding(
tmp_path, empty_lerobot_dataset_factory, monkeypatch, dtype, np_dtype, values, assert_fn
):
features = {"state": {"dtype": dtype, "shape": (1,), "names": None}}
dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)
dataset.add_frame({"state": np.array([values[0]], dtype=np_dtype), "task": "Dummy task"})
dataset.add_frame({"state": np.array([values[1]], dtype=np_dtype), "task": "Dummy task"})
captured = {}
original_from_dict = datasets.Dataset.from_dict
def _from_dict_spy(cls, mapping, *args, **kwargs):
captured["state"] = mapping["state"]
return original_from_dict(mapping, *args, **kwargs)
monkeypatch.setattr(datasets.Dataset, "from_dict", classmethod(_from_dict_spy))
dataset.save_episode()
dataset.finalize()
assert "state" in captured
assert isinstance(captured["state"], np.ndarray)
assert captured["state"].shape == (2,)
assert_fn(captured["state"], np.array(values, dtype=np_dtype))
def test_set_image_transforms_applies_transparently(image_dataset):
dataset = image_dataset
dataset.add_frame({"image": np.random.rand(*DUMMY_CHW), "task": "Dummy task"})
@@ -1470,17 +1506,9 @@ def test_frames_in_current_file_calculation(tmp_path, empty_lerobot_dataset_fact
def test_lerobot_dataset_vcodec_validation():
"""Test that LeRobotDataset validates the vcodec parameter."""
# Test that invalid vcodec raises ValueError
"""Invalid vcodec in encoder config is rejected at construction time."""
with pytest.raises(ValueError, match="Invalid vcodec"):
LeRobotDataset.__new__(LeRobotDataset) # bypass __init__ to test validation directly
# Actually test via create since it's easier
LeRobotDataset.create(
repo_id="test/invalid_codec",
fps=30,
features={"observation.state": {"dtype": "float32", "shape": (2,), "names": ["x", "y"]}},
vcodec="invalid_codec",
)
VideoEncoderConfig(vcodec="invalid_codec")
def test_valid_video_codecs_constant():
tests/datasets/test_language.py
+173
View File
@@ -0,0 +1,173 @@
#!/usr/bin/env python
import pytest
pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
pytest.importorskip("pandas", reason="pandas is required (install lerobot[dataset])")
import numpy as np # noqa: E402
import pandas as pd # noqa: E402
import pyarrow as pa # noqa: E402
from lerobot.datasets import LeRobotDataset # noqa: E402
from lerobot.datasets.io_utils import write_info # noqa: E402
from lerobot.datasets.language import ( # noqa: E402
EVENT_ONLY_STYLES,
LANGUAGE_EVENTS,
LANGUAGE_PERSISTENT,
PERSISTENT_STYLES,
STYLE_REGISTRY,
VIEW_DEPENDENT_STYLES,
column_for_style,
is_view_dependent_style,
language_events_arrow_type,
language_feature_info,
language_persistent_arrow_type,
validate_camera_field,
)
from lerobot.datasets.utils import DEFAULT_DATA_PATH # noqa: E402
def test_language_arrow_schema_has_expected_fields():
persistent_row_type = language_persistent_arrow_type().value_type
event_row_type = language_events_arrow_type().value_type
assert isinstance(persistent_row_type, pa.StructType)
assert persistent_row_type.names == [
"role",
"content",
"style",
"timestamp",
"camera",
"tool_calls",
]
assert isinstance(event_row_type, pa.StructType)
assert event_row_type.names == ["role", "content", "style", "camera", "tool_calls"]
# Persistent-row timestamps use float32, matching LeRobotDataset frame timestamps.
assert persistent_row_type.field("timestamp").type == pa.float32()
def test_validate_feature_language_warns_only_on_non_empty_value(caplog):
from lerobot.datasets.feature_utils import validate_feature_language
# None (the expected record-time value) is silent and non-fatal.
with caplog.at_level("WARNING"):
assert validate_feature_language("language_persistent", None) == ""
assert caplog.records == []
# A stray non-empty value is dropped later, so we warn rather than fail.
with caplog.at_level("WARNING"):
assert validate_feature_language("language_persistent", [{"role": "user"}]) == ""
assert any("language_persistent" in r.message for r in caplog.records)
def test_style_registry_routes_columns():
assert {"subtask", "plan", "memory", "motion", "task_aug"} == PERSISTENT_STYLES
assert {"interjection", "vqa", "trace"} == EVENT_ONLY_STYLES
assert PERSISTENT_STYLES | EVENT_ONLY_STYLES <= STYLE_REGISTRY
assert column_for_style("subtask") == LANGUAGE_PERSISTENT
assert column_for_style("plan") == LANGUAGE_PERSISTENT
assert column_for_style("memory") == LANGUAGE_PERSISTENT
assert column_for_style("motion") == LANGUAGE_PERSISTENT
assert column_for_style("task_aug") == LANGUAGE_PERSISTENT
assert column_for_style("interjection") == LANGUAGE_EVENTS
assert column_for_style("vqa") == LANGUAGE_EVENTS
assert column_for_style("trace") == LANGUAGE_EVENTS
assert column_for_style(None) == LANGUAGE_EVENTS
def test_view_dependent_styles():
# motion lives in PERSISTENT_STYLES and is described in robot-frame
# (joint / Cartesian) terms, so it is NOT view-dependent. Only vqa
# (event) and trace (event, pixel-trajectory) carry a camera tag.
assert {"vqa", "trace"} == VIEW_DEPENDENT_STYLES
assert is_view_dependent_style("vqa")
assert is_view_dependent_style("trace")
assert not is_view_dependent_style("motion")
assert not is_view_dependent_style("subtask")
assert not is_view_dependent_style("plan")
assert not is_view_dependent_style("interjection")
assert not is_view_dependent_style(None)
def test_validate_camera_field_requires_camera_for_view_dependent_styles():
validate_camera_field("vqa", "observation.images.top")
validate_camera_field("trace", "observation.images.front")
with pytest.raises(ValueError, match="view-dependent"):
validate_camera_field("vqa", None)
with pytest.raises(ValueError, match="view-dependent"):
validate_camera_field("trace", "")
def test_validate_camera_field_rejects_camera_on_non_view_dependent_styles():
validate_camera_field("subtask", None)
validate_camera_field("plan", None)
validate_camera_field("memory", None)
validate_camera_field("motion", None)
validate_camera_field("interjection", None)
validate_camera_field(None, None)
with pytest.raises(ValueError, match="must have camera=None"):
validate_camera_field("subtask", "observation.images.top")
with pytest.raises(ValueError, match="must have camera=None"):
validate_camera_field("motion", "observation.images.top")
with pytest.raises(ValueError, match="must have camera=None"):
validate_camera_field("interjection", "observation.images.top")
with pytest.raises(ValueError, match="must have camera=None"):
validate_camera_field(None, "observation.images.top")
def test_unknown_style_rejected():
with pytest.raises(ValueError, match="Unknown language style"):
column_for_style("surprise")
def test_lerobot_dataset_passes_language_columns_through(tmp_path, empty_lerobot_dataset_factory):
root = tmp_path / "language_dataset"
dataset = empty_lerobot_dataset_factory(
root=root,
features={"state": {"dtype": "float32", "shape": (2,), "names": None}},
use_videos=False,
)
dataset.add_frame({"state": np.array([0.0, 1.0], dtype=np.float32), "task": "tidy"})
dataset.add_frame({"state": np.array([1.0, 2.0], dtype=np.float32), "task": "tidy"})
dataset.save_episode()
dataset.finalize()
persistent = [
{
"role": "assistant",
"content": "reach for the cup",
"style": "subtask",
"timestamp": 0.0,
"camera": None,
"tool_calls": None,
}
]
event = {
"role": "user",
"content": "what is visible?",
"style": "vqa",
"camera": "observation.images.top",
"tool_calls": None,
}
data_path = root / DEFAULT_DATA_PATH.format(chunk_index=0, file_index=0)
df = pd.read_parquet(data_path)
df[LANGUAGE_PERSISTENT] = [persistent, persistent]
df[LANGUAGE_EVENTS] = [[event], []]
df.to_parquet(data_path)
info = dataset.meta.info
info["features"].update(language_feature_info())
write_info(info, root)
reloaded = LeRobotDataset(repo_id=dataset.repo_id, root=root)
first = reloaded[0]
second = reloaded[1]
assert first[LANGUAGE_PERSISTENT] == persistent
assert first[LANGUAGE_EVENTS] == [event]
assert second[LANGUAGE_PERSISTENT] == persistent
assert second[LANGUAGE_EVENTS] == []
tests/datasets/test_language_render.py
+417
View File
@@ -0,0 +1,417 @@
#!/usr/bin/env python
import pytest
pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
from lerobot.configs.recipe import MessageTurn, TrainingRecipe # noqa: E402
from lerobot.datasets.language_render import ( # noqa: E402
EMITTED_AT_TOLERANCE_S,
active_at,
emitted_at,
nth_next,
nth_prev,
render_sample,
)
def persistent_row(role, content, style, timestamp, tool_calls=None, camera=None):
return {
"role": role,
"content": content,
"style": style,
"timestamp": timestamp,
"camera": camera,
"tool_calls": tool_calls,
}
def event_row(role, content, style, tool_calls=None, camera=None):
return {
"role": role,
"content": content,
"style": style,
"camera": camera,
"tool_calls": tool_calls,
}
PERSISTENT = [
persistent_row("assistant", "plan 0", "plan", 0.0),
persistent_row("assistant", "memory 0", "memory", 0.0),
persistent_row("assistant", "subtask 0", "subtask", 0.0),
persistent_row("assistant", "memory 1", "memory", 1.0),
persistent_row("assistant", "subtask 1", "subtask", 1.0),
]
EVENTS_AT_1 = [
event_row("user", "what is visible?", "vqa", camera="observation.images.top"),
event_row("assistant", '{"count": 2}', "vqa", camera="observation.images.top"),
]
EVENTS_AT_2 = [
event_row("user", "skip wiping", "interjection"),
event_row(
"assistant",
None,
None,
[{"type": "function", "function": {"name": "say", "arguments": {"text": "Skipping wiping."}}}],
),
]
# Same emission tick, two cameras: triggers per-camera disambiguation in
# resolvers, mirroring how Module 3 of the annotation pipeline writes one
# (vqa, user) + (vqa, assistant) pair per camera.
EVENTS_AT_3_TWO_CAMERAS = [
event_row("user", "how many cups (top)?", "vqa", camera="observation.images.top"),
event_row("assistant", '{"count": 3}', "vqa", camera="observation.images.top"),
event_row("user", "how many cups (wrist)?", "vqa", camera="observation.images.wrist"),
event_row("assistant", '{"count": 1}', "vqa", camera="observation.images.wrist"),
]
def test_resolver_temporal_semantics():
assert active_at(0.5, persistent=PERSISTENT, style="subtask")["content"] == "subtask 0"
assert active_at(1.0, persistent=PERSISTENT, style="subtask")["content"] == "subtask 1"
assert emitted_at(0.5, persistent=PERSISTENT, events=[], style="vqa", role="assistant") is None
assert (
emitted_at(1.0, persistent=PERSISTENT, events=EVENTS_AT_1, style="vqa", role="assistant")["content"]
== '{"count": 2}'
)
def test_persistent_relative_resolvers_reject_event_styles():
with pytest.raises(ValueError, match="event-only"):
active_at(1.0, persistent=PERSISTENT, style="vqa")
with pytest.raises(ValueError, match="event-only"):
nth_prev(1.0, persistent=PERSISTENT, style="interjection")
def test_nth_prev_and_next():
assert nth_prev(1.0, persistent=PERSISTENT, style="subtask", offset=1)["content"] == "subtask 0"
assert nth_next(0.0, persistent=PERSISTENT, style="subtask", offset=1)["content"] == "subtask 1"
def test_substitution_if_present_multimodal_and_tool_calls():
recipe = TrainingRecipe(
messages=[
MessageTurn(
role="user",
content=[
{"type": "image", "feature": "observation.images.top"},
{"type": "text", "text": "${task}: ${interjection}"},
],
stream="high_level",
if_present="interjection",
),
MessageTurn(
role="assistant",
content="${plan}",
stream="high_level",
target=True,
tool_calls_from="speech",
),
],
bindings={"plan": "active_at(t, style=plan)"},
)
rendered = render_sample(
recipe=recipe,
persistent=PERSISTENT,
events=EVENTS_AT_2,
t=2.0,
sample_idx=0,
task="clean kitchen",
)
assert rendered["messages"][0]["content"][1]["text"] == "clean kitchen: skip wiping"
assert rendered["messages"][1]["content"] == "plan 0"
assert rendered["messages"][1]["tool_calls"][0]["function"]["name"] == "say"
assert rendered["message_streams"] == ["high_level", "high_level"]
assert rendered["target_message_indices"] == [1]
def test_exact_event_miss_returns_none_when_target_skips():
recipe = TrainingRecipe(
messages=[
MessageTurn(role="user", content="${vqa_query}", stream="high_level", if_present="vqa_query"),
MessageTurn(
role="assistant",
content="${vqa}",
stream="high_level",
target=True,
if_present="vqa",
),
]
)
assert (
render_sample(recipe=recipe, persistent=PERSISTENT, events=EVENTS_AT_2, t=0.0, sample_idx=0) is None
)
def test_deterministic_blend_sampling():
recipe = TrainingRecipe(
blend={
"a": TrainingRecipe(
weight=1.0,
messages=[
MessageTurn(role="user", content="${task}", stream="high_level"),
MessageTurn(role="assistant", content="a", stream="high_level", target=True),
],
),
"b": TrainingRecipe(
weight=1.0,
messages=[
MessageTurn(role="user", content="${task}", stream="high_level"),
MessageTurn(role="assistant", content="b", stream="high_level", target=True),
],
),
}
)
first = render_sample(
recipe=recipe, persistent=PERSISTENT, events=EVENTS_AT_2, t=0.0, sample_idx=123, task="x"
)
second = render_sample(
recipe=recipe, persistent=PERSISTENT, events=EVENTS_AT_2, t=0.0, sample_idx=123, task="x"
)
assert first == second
def test_emitted_at_filters_vqa_by_camera():
top = emitted_at(
3.0,
persistent=PERSISTENT,
events=EVENTS_AT_3_TWO_CAMERAS,
style="vqa",
role="assistant",
camera="observation.images.top",
)
wrist = emitted_at(
3.0,
persistent=PERSISTENT,
events=EVENTS_AT_3_TWO_CAMERAS,
style="vqa",
role="assistant",
camera="observation.images.wrist",
)
assert top["content"] == '{"count": 3}'
assert wrist["content"] == '{"count": 1}'
def test_emitted_at_raises_on_ambiguous_per_camera_vqa():
with pytest.raises(ValueError, match="Ambiguous resolver"):
emitted_at(
3.0,
persistent=PERSISTENT,
events=EVENTS_AT_3_TWO_CAMERAS,
style="vqa",
role="assistant",
)
def _vqa_subrecipe(camera: str) -> TrainingRecipe:
return TrainingRecipe(
weight=1.0,
bindings={
"vqa_query": f"emitted_at(t, style=vqa, role=user, camera={camera})",
"vqa": f"emitted_at(t, style=vqa, role=assistant, camera={camera})",
},
messages=[
MessageTurn(
role="user",
content=[{"type": "image", "feature": camera}, {"type": "text", "text": "${vqa_query}"}],
stream="high_level",
if_present="vqa_query",
),
MessageTurn(
role="assistant",
content="${vqa}",
stream="high_level",
target=True,
if_present="vqa",
),
],
)
@pytest.mark.parametrize(
("camera", "expected_query", "expected_answer"),
[
("observation.images.top", "how many cups (top)?", '{"count": 3}'),
("observation.images.wrist", "how many cups (wrist)?", '{"count": 1}'),
],
)
def test_per_camera_blend_renders_both_views(camera, expected_query, expected_answer):
rendered = render_sample(
recipe=_vqa_subrecipe(camera),
persistent=PERSISTENT,
events=EVENTS_AT_3_TWO_CAMERAS,
t=3.0,
sample_idx=0,
)
assert rendered["messages"][0]["content"][0]["feature"] == camera
assert rendered["messages"][0]["content"][1]["text"] == expected_query
assert rendered["messages"][1]["content"] == expected_answer
def test_resolve_task_picks_rephrasing_deterministically_per_sample():
rephrasings = [
persistent_row("user", "tidy the kitchen", "task_aug", 0.0),
persistent_row("user", "please clean up the kitchen", "task_aug", 0.0),
persistent_row("user", "kitchen needs tidying", "task_aug", 0.0),
persistent_row("user", "make the kitchen clean", "task_aug", 0.0),
]
recipe = TrainingRecipe(
messages=[
MessageTurn(role="user", content="${task}", stream="high_level"),
MessageTurn(role="assistant", content="ok", stream="high_level", target=True),
]
)
# No explicit task override → resolver consults persistent rows.
seen: set[str] = set()
for sample_idx in range(64):
rendered = render_sample(
recipe=recipe,
persistent=rephrasings,
events=[],
t=0.0,
sample_idx=sample_idx,
dataset_ctx={"task": "canonical kitchen task"},
)
seen.add(rendered["messages"][0]["content"])
# Every rephrasing should be reachable across enough samples.
assert seen == {r["content"] for r in rephrasings}
# Same sample_idx → same pick (determinism).
a = render_sample(
recipe=recipe,
persistent=rephrasings,
events=[],
t=0.0,
sample_idx=42,
dataset_ctx={"task": "canonical"},
)
b = render_sample(
recipe=recipe,
persistent=rephrasings,
events=[],
t=0.0,
sample_idx=42,
dataset_ctx={"task": "canonical"},
)
assert a["messages"][0]["content"] == b["messages"][0]["content"]
def test_resolve_task_falls_back_to_canonical_without_rephrasings():
recipe = TrainingRecipe(
messages=[
MessageTurn(role="user", content="${task}", stream="high_level"),
MessageTurn(role="assistant", content="ok", stream="high_level", target=True),
]
)
rendered = render_sample(
recipe=recipe,
persistent=PERSISTENT, # no task_aug rows
events=[],
t=0.0,
sample_idx=0,
dataset_ctx={"task": "clean the kitchen"},
)
assert rendered["messages"][0]["content"] == "clean the kitchen"
def test_resolve_task_explicit_override_beats_rephrasings():
rephrasings = [
persistent_row("user", "rephrased one", "task_aug", 0.0),
persistent_row("user", "rephrased two", "task_aug", 0.0),
]
recipe = TrainingRecipe(
messages=[
MessageTurn(role="user", content="${task}", stream="high_level"),
MessageTurn(role="assistant", content="ok", stream="high_level", target=True),
]
)
rendered = render_sample(
recipe=recipe,
persistent=rephrasings,
events=[],
t=0.0,
sample_idx=0,
task="explicit override wins",
dataset_ctx={"task": "canonical"},
)
assert rendered["messages"][0]["content"] == "explicit override wins"
def test_emitted_at_persistent_tolerates_small_timestamp_drift():
"""Persistent ``emitted_at`` should match within EMITTED_AT_TOLERANCE_S
so callers that derive ``t`` arithmetically (``frame_idx / fps``) still
line up with the parquet-stored timestamp.
"""
rows = [persistent_row("assistant", "memo", "memory", 1.0)]
# Half a tolerance window — bit-different float, comfortably inside
inside = emitted_at(1.0 + EMITTED_AT_TOLERANCE_S / 2, persistent=rows, events=[], style="memory")
assert inside is not None and inside["content"] == "memo"
# Just past the window — no match
outside = emitted_at(1.0 + EMITTED_AT_TOLERANCE_S * 2, persistent=rows, events=[], style="memory")
assert outside is None
def test_render_sample_rejects_non_dict_language_rows():
"""``_normalize_rows`` must surface malformed inputs as TypeError.
A pipeline that hands the renderer a non-dict (e.g. a stray string)
is a real upstream bug — silent skipping would let it propagate.
"""
recipe = TrainingRecipe(
messages=[
MessageTurn(role="user", content="${task}", stream="high_level"),
MessageTurn(role="assistant", content="ok", stream="high_level", target=True),
]
)
with pytest.raises(TypeError, match="must be dictionaries"):
render_sample(
recipe=recipe,
persistent=["not a dict"],
events=[],
t=0.0,
sample_idx=0,
task="x",
)
def test_low_level_branch_renders_active_subtask():
low_level = TrainingRecipe(
blend={
"low": TrainingRecipe(
weight=1.0,
messages=[
MessageTurn(
role="user",
content="${task}\nPlan: ${plan}\nMemory: ${memory}",
stream="high_level",
),
MessageTurn(
role="assistant",
content="${subtask}",
stream="low_level",
target=True,
),
],
)
}
)
rendered = render_sample(
recipe=low_level,
persistent=PERSISTENT,
events=[],
t=0.5,
sample_idx=0,
task="clean kitchen",
)
assert rendered["messages"][-1] == {"role": "assistant", "content": "subtask 0"}
assert rendered["message_streams"][-1] == "low_level"
assert rendered["target_message_indices"] == [1]

Some files were not shown because too many files have changed in this diff Show More