feat(training): add inline offline validation with train/eval split

- Add eval_split config for balanced per-task holdout
- Add eval_steps for periodic inline eval loss computation
- Add max_eval_samples to cap eval cost

.github/workflows/benchmark_tests.yml

@@ -167,9 +167,9 @@ jobs:
 
       # ── LIBERO TRAIN+EVAL SMOKE ──────────────────────────────────────────────
       # Train SmolVLA for 1 step (batch_size=1, dataset episode 0 only) then
-      # immediately runs eval inside the training loop (eval_freq=1, 1 episode).
+      # immediately runs eval inside the training loop (env_eval_freq=1, 1 episode).
       # Tests the full train→eval-within-training pipeline end-to-end.
-      - name: Run Libero train+eval smoke (1 step, eval_freq=1)
+      - name: Run Libero train+eval smoke (1 step, env_eval_freq=1)
         if: env.HF_USER_TOKEN != ''
         run: |
           docker run --name libero-train-smoke --gpus all \
@@ -196,7 +196,7 @@ jobs:
                 --output_dir=/tmp/train-smoke \
                 --steps=1 \
                 --batch_size=1 \
-                --eval_freq=1 \
+                --env_eval_freq=1 \
                 --eval.n_episodes=1 \
                 --eval.batch_size=1 \
                 --eval.use_async_envs=false \

Makefile

@@ -58,7 +58,7 @@ test-act-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=4 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_freq=2 \
@@ -96,7 +96,7 @@ test-diffusion-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=2 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_checkpoint=true \
@@ -126,7 +126,7 @@ test-tdmpc-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=2 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_checkpoint=true \
@@ -161,7 +161,7 @@ test-smolvla-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=4 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_freq=2 \

docs/source/hil_data_collection.mdx

@@ -57,11 +57,11 @@ The `lerobot-rollout --strategy.type=dagger` mode requires **teleoperators with
 
 **Compatible teleoperators:**
 
-- `bi_openarm_mini` - Bimanual OpenArm Mini
+- `openarm_mini` - OpenArm Mini
 - `so_leader` - SO100 / SO101 leader arm
 
 > [!IMPORTANT]
-> The provided commands default to `bi_openarm_follower` + `bi_openarm_mini`.
+> The provided commands default to `bi_openarm_follower` + `openarm_mini`.
 > `so_follower` + `so_leader` configs are also registered and can be used via CLI flags.
 
 ---
@@ -104,9 +104,9 @@ lerobot-rollout --strategy.type=dagger \
     --robot.right_arm_config.port=can0 \
     --robot.right_arm_config.side=right \
     --robot.cameras='{left_wrist: {type: opencv, index_or_path: "/dev/video0", width: 1280, height: 720, fps: 30}, right_wrist: {type: opencv, index_or_path: "/dev/video4", width: 1280, height: 720, fps: 30}, base: {type: opencv, index_or_path: "/dev/video2", width: 640, height: 480, fps: 30}}' \
-    --teleop.type=bi_openarm_mini \
-    --teleop.left_arm_config.port=/dev/ttyACM0 \
-    --teleop.right_arm_config.port=/dev/ttyACM1 \
+    --teleop.type=openarm_mini \
+    --teleop.port_left=/dev/ttyACM0 \
+    --teleop.port_right=/dev/ttyACM1 \
     --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
     --dataset.repo_id=your-username/rollout_hil_dataset \
     --dataset.single_task="Fold the T-shirt properly" \
@@ -131,9 +131,9 @@ lerobot-rollout --strategy.type=dagger \
     --robot.right_arm_config.port=can0 \
     --robot.right_arm_config.side=right \
     --robot.cameras='{left_wrist: {type: opencv, index_or_path: "/dev/video0", width: 1280, height: 720, fps: 30}, right_wrist: {type: opencv, index_or_path: "/dev/video4", width: 1280, height: 720, fps: 30}, base: {type: opencv, index_or_path: "/dev/video2", width: 640, height: 480, fps: 30}}' \
-    --teleop.type=bi_openarm_mini \
-    --teleop.left_arm_config.port=/dev/ttyACM0 \
-    --teleop.right_arm_config.port=/dev/ttyACM1 \
+    --teleop.type=openarm_mini \
+    --teleop.port_left=/dev/ttyACM0 \
+    --teleop.port_right=/dev/ttyACM1 \
     --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
     --dataset.repo_id=your-username/rollout_hil_rtc_dataset \
     --dataset.single_task="Fold the T-shirt properly" \

docs/source/hilserl.mdx

@@ -719,7 +719,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
   "num_workers": 4,
   "steps": 5000,
   "log_freq": 10,
-  "eval_freq": 1000,
+  "env_eval_freq": 1000,
   "save_freq": 1000,
   "save_checkpoint": true,
   "seed": 2,

docs/source/inference.mdx

@@ -117,7 +117,7 @@ lerobot-rollout \
     --strategy.num_episodes=20 \
     --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
     --robot.type=bi_openarm_follower \
-    --teleop.type=bi_openarm_mini \
+    --teleop.type=openarm_mini \
     --dataset.repo_id=${HF_USER}/rollout_hil_data \
     --dataset.single_task="Fold the T-shirt"
 ```

docs/source/libero.mdx

@@ -143,7 +143,7 @@ lerobot-train \
   --batch_size=4 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
-  --eval_freq=1000
+  --env_eval_freq=1000
 ```
 
 ## Reproducing published results

docs/source/libero_plus.mdx

@@ -173,7 +173,7 @@ lerobot-train \
     --batch_size=4 \
     --eval.batch_size=1 \
     --eval.n_episodes=1 \
-    --eval_freq=1000
+    --env_eval_freq=1000
 ```
 
 ## Relationship to LIBERO

docs/source/metaworld.mdx

@@ -120,11 +120,11 @@ lerobot-train \
   --batch_size=4 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
-  --eval_freq=1000
+  --env_eval_freq=1000
 ```
 
 ## Practical tips
 
 - Use the one-hot task conditioning for multi-task training (MT10/MT50 conventions) so policies have explicit task context.
 - Inspect the dataset task descriptions and the `info["is_success"]` keys when writing post-processing or logging so your success metrics line up with the benchmark.
-- Adjust `batch_size`, `steps`, and `eval_freq` to match your compute budget.
+- Adjust `batch_size`, `steps`, and `env_eval_freq` to match your compute budget.

docs/source/molmoact2.mdx

@@ -103,7 +103,7 @@ accelerate launch \
   --batch_size=32 \
   --num_workers=4 \
   --log_freq=20 \
-  --eval_freq=-1 \
+  --env_eval_freq=-1 \
   --save_checkpoint=true \
   --save_freq=2000
 ```
@@ -142,7 +142,7 @@ accelerate launch \
   --batch_size=32 \
   --num_workers=4 \
   --log_freq=20 \
-  --eval_freq=-1 \
+  --env_eval_freq=-1 \
   --save_checkpoint=true \
   --save_freq=2000
 ```

docs/source/multi_task_dit.mdx

@@ -314,7 +314,7 @@ lerobot-train \
   --steps=30000 \
   --save_freq=1000 \
   --log_freq=100 \
-  --eval_freq=1000 \
+  --env_eval_freq=1000 \
   --policy.type=multi_task_dit \
   --policy.device=cuda \
   --policy.horizon=32 \

docs/source/robocasa.mdx

@@ -166,7 +166,7 @@ lerobot-train \
   --output_dir=./outputs/smolvla_robocasa_CloseFridge \
   --steps=100000 \
   --batch_size=4 \
-  --eval_freq=5000 \
+  --env_eval_freq=5000 \
   --eval.batch_size=1 \
   --eval.n_episodes=5 \
   --save_freq=10000

docs/source/vlabench.mdx

@@ -165,7 +165,7 @@ lerobot-train \
   --output_dir=./outputs/smolvla_vlabench_primitive \
   --steps=100000 \
   --batch_size=4 \
-  --eval_freq=5000 \
+  --env_eval_freq=5000 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
   --save_freq=10000

src/lerobot/annotations/steerable_pipeline/writer.py

@@ -54,7 +54,6 @@ from typing import Any
 import pyarrow as pa
 import pyarrow.parquet as pq
 
-from lerobot.datasets.io_utils import write_table_one_row_group_per_episode
 from lerobot.datasets.language import (
     EVENT_ONLY_STYLES,
     LANGUAGE_EVENTS,
@@ -275,11 +274,12 @@ class LanguageColumnsWriter:
         new_table = self._materialize_table(
             table, per_row_persistent, per_row_events, drop_old=self.drop_existing_subtask_index
         )
-        # Re-emit one row group per episode (a bulk pq.write_table would collapse
-        # them into one). Write to a sibling tmp path and atomically rename so a
-        # crash mid-write can't leave a half-written shard.
+        # Atomic replace: write to a sibling tmp path and rename so a crash
+        # mid-write can't leave a half-written shard that ``pq.read_table``
+        # would then fail to open. ``Path.replace`` is atomic on POSIX +
+        # Windows when source and target sit on the same filesystem.
         tmp_path = path.with_suffix(path.suffix + ".tmp")
-        write_table_one_row_group_per_episode(new_table, tmp_path)
+        pq.write_table(new_table, tmp_path)
         tmp_path.replace(path)
 
     def _materialize_table(

src/lerobot/configs/default.py

@@ -39,6 +39,8 @@ class DatasetConfig:
     # This reduces memory and speeds up DataLoader IPC. The training pipeline handles the conversion.
     return_uint8: bool = False
     streaming: bool = False
+    # Fraction of episodes held out per task for offline evaluation (0.0 = disabled).
+    eval_split: float = 0.0
 
     def __post_init__(self) -> None:
         if self.episodes is not None:

src/lerobot/configs/train.py

@@ -100,8 +100,13 @@ class TrainPipelineConfig(HubMixin):
     prefetch_factor: int = 4
     persistent_workers: bool = True
     steps: int = 100_000
-    eval_freq: int = 20_000
+    # Run policy in the simulation environment every N steps to measure reward/success (0 = disabled).
+    env_eval_freq: int = 20_000
     log_freq: int = 200
+    # Compute eval loss on held-out episodes every N steps (0 = disabled). Requires eval_split > 0.
+    eval_steps: int = 0
+    # Cap on total eval samples, split uniformly across tasks (0 = use all held-out data).
+    max_eval_samples: int = 0
     tolerance_s: float = 1e-4
     save_checkpoint: bool = True
     # Checkpoint is saved every `save_freq` training iterations and after the last training step.

src/lerobot/datasets/__init__.py

@@ -35,7 +35,7 @@ from .dataset_tools import (
     remove_feature,
     split_dataset,
 )
-from .factory import make_dataset, resolve_delta_timestamps
+from .factory import make_dataset, make_train_eval_datasets, resolve_delta_timestamps
 from .image_writer import safe_stop_image_writer
 from .io_utils import load_episodes, write_stats
 from .language import (
@@ -89,6 +89,7 @@ __all__ = [
     "get_feature_stats",
     "load_episodes",
     "make_dataset",
+    "make_train_eval_datasets",
     "merge_datasets",
     "modify_features",
     "modify_tasks",

src/lerobot/datasets/aggregate.py

@@ -32,7 +32,6 @@ from .feature_utils import features_equal_for_merge, get_hf_features_from_featur
 from .io_utils import (
     get_file_size_in_mb,
     get_parquet_file_size_in_mb,
-    to_parquet_one_row_group_per_episode,
     to_parquet_with_hf_images,
     write_info,
     write_stats,
@@ -552,7 +551,6 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
             aggr_root=dst_meta.root,
             hf_features=hf_features,
             concatenate=concatenate_data,
-            one_row_group_per_episode=True,
         )
 
         # Record the mapping from source to actual destination
@@ -630,7 +628,6 @@ def append_or_create_parquet_file(
     aggr_root: Path = None,
     hf_features: datasets.Features | None = None,
     concatenate: bool = True,
-    one_row_group_per_episode: bool = False,
 ) -> tuple[dict[str, int], tuple[int, int]]:
     """Appends data to an existing parquet file or creates a new one based on size constraints.
 
@@ -648,8 +645,6 @@ def append_or_create_parquet_file(
         aggr_root: Root path for the aggregated dataset.
         hf_features: Optional HuggingFace Features schema for proper image typing.
         concatenate: When False, always rotate to a new file instead of appending to the current one.
-        one_row_group_per_episode: True for DATA parquet (emit one row group per episode); False for
-            the episodes-metadata parquet (already one row per episode).
 
     Returns:
         tuple: (updated_idx, (dst_chunk, dst_file)) where updated_idx is the index dict
@@ -662,8 +657,6 @@ def append_or_create_parquet_file(
         dst_path.parent.mkdir(parents=True, exist_ok=True)
         if contains_images:
             to_parquet_with_hf_images(df, dst_path, features=hf_features)
-        elif one_row_group_per_episode:
-            to_parquet_one_row_group_per_episode(df, dst_path)
         else:
             df.to_parquet(dst_path)
         return idx, (dst_chunk, dst_file)
@@ -690,8 +683,6 @@ def append_or_create_parquet_file(
 
     if contains_images:
         to_parquet_with_hf_images(final_df, target_path, features=hf_features)
-    elif one_row_group_per_episode:
-        to_parquet_one_row_group_per_episode(final_df, target_path)
     else:
         final_df.to_parquet(target_path)
 

src/lerobot/datasets/factory.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import logging
+import math
 from pprint import pformat
 
 import torch
@@ -130,3 +131,81 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
                 dataset.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
 
     return dataset
+
+
+def make_train_eval_datasets(
+    cfg: TrainPipelineConfig,
+) -> tuple[LeRobotDataset | MultiLeRobotDataset, LeRobotDataset | None]:
+    """Create train and optional eval datasets by splitting episodes based on eval_split.
+
+    The last ceil(n_episodes * eval_split) episodes per task are held out for evaluation.
+    If eval_split == 0.0, returns (full_dataset, None).
+    """
+    full_dataset = make_dataset(cfg)
+
+    if cfg.dataset.eval_split == 0.0:
+        return full_dataset, None
+
+    base_episodes = (
+        full_dataset.episodes if full_dataset.episodes is not None else list(range(full_dataset.num_episodes))
+    )
+
+    episode_tasks = full_dataset.meta.episodes["tasks"]
+    task_to_episodes: dict[str, list[int]] = {}
+    for ep_idx in base_episodes:
+        task_key = episode_tasks[ep_idx][0] if episode_tasks[ep_idx] else ""
+        task_to_episodes.setdefault(task_key, []).append(ep_idx)
+
+    train_episodes, eval_episodes = [], []
+    for eps in task_to_episodes.values():
+        n_eval = math.ceil(len(eps) * cfg.dataset.eval_split)
+        train_episodes.extend(eps[: len(eps) - n_eval])
+        eval_episodes.extend(eps[len(eps) - n_eval :])
+
+    if not train_episodes:
+        raise ValueError(
+            f"eval_split={cfg.dataset.eval_split} leaves 0 training episodes from {len(base_episodes)} total."
+        )
+
+    logging.info(
+        f"Train/eval split: {len(train_episodes)} train, {len(eval_episodes)} eval "
+        f"(eval_split={cfg.dataset.eval_split}, {len(task_to_episodes)} tasks)"
+    )
+
+    delta_timestamps = resolve_delta_timestamps(cfg.trainable_config, full_dataset.meta)
+
+    train_image_transforms = (
+        ImageTransforms(cfg.dataset.image_transforms) if cfg.dataset.image_transforms.enable else None
+    )
+
+    train_dataset = LeRobotDataset(
+        cfg.dataset.repo_id,
+        root=cfg.dataset.root,
+        episodes=train_episodes,
+        delta_timestamps=delta_timestamps,
+        image_transforms=train_image_transforms,
+        revision=cfg.dataset.revision,
+        video_backend=cfg.dataset.video_backend,
+        return_uint8=True,
+        tolerance_s=cfg.tolerance_s,
+    )
+
+    eval_dataset = LeRobotDataset(
+        cfg.dataset.repo_id,
+        root=cfg.dataset.root,
+        episodes=eval_episodes,
+        delta_timestamps=delta_timestamps,
+        image_transforms=None,
+        revision=cfg.dataset.revision,
+        video_backend=cfg.dataset.video_backend,
+        return_uint8=True,
+        tolerance_s=cfg.tolerance_s,
+    )
+
+    if cfg.dataset.use_imagenet_stats:
+        for ds in (train_dataset, eval_dataset):
+            for key in ds.meta.camera_keys:
+                for stats_type, stats in IMAGENET_STATS.items():
+                    ds.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
+
+    return train_dataset, eval_dataset

src/lerobot/datasets/io_utils.py

@@ -20,7 +20,6 @@ import datasets
 import numpy as np
 import pandas
 import pandas as pd
-import pyarrow as pa
 import pyarrow.dataset as pa_ds
 import pyarrow.parquet as pq
 import torch
@@ -271,49 +270,21 @@ def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[to
     return items_dict
 
 
-def write_table_one_row_group_per_episode(table: pa.Table, path: Path) -> None:
-    """Write ``table`` with one parquet row group per episode (in episode order).
-
-    Keeps shards random-access friendly (``read_row_group(i)`` fetches episode i),
-    mirroring the recording writer. ``table`` must carry a contiguous
-    ``episode_index`` column.
-    """
-        episode_index = table.column("episode_index").to_numpy(zero_copy_only=False)
-    starts = np.concatenate(([0], np.nonzero(np.diff(episode_index))[0] + 1))
-    writer = pq.ParquetWriter(str(path), table.schema, compression="snappy", use_dictionary=True)
-    try:
-        for start, stop in zip(starts, np.append(starts[1:], len(episode_index))):
-            writer.write_table(table.slice(start, stop - start))  # one episode -> one row group
-    finally:
-        writer.close()
-
-
 def to_parquet_with_hf_images(
     df: pandas.DataFrame, path: Path, features: datasets.Features | None = None
 ) -> None:
-    """Write a DataFrame with HF-encoded images to parquet, one row group per episode.
+    """This function correctly writes to parquet a panda DataFrame that contains images encoded by HF dataset.
+    This way, it can be loaded by HF dataset and correctly formatted images are returned.
 
-    Images are embedded into the arrow table first (``ParquetWriter.write_table``
-    does not embed external image files like ``Dataset.to_parquet`` does).
-    ``features`` types image columns as ``Image()`` in the parquet schema.
+    Args:
+        df: DataFrame to write to parquet.
+        path: Path to write the parquet file.
+        features: Optional HuggingFace Features schema. If provided, ensures image columns
+                  are properly typed as Image() in the parquet schema.
     """
+    # TODO(qlhoest): replace this weird synthax by `df.to_parquet(path)` only
     ds = datasets.Dataset.from_dict(df.to_dict(orient="list"), features=features)
-    ds = embed_images(ds)
-    table = ds.with_format("arrow")[:]
-    if "episode_index" in table.column_names:
-        write_table_one_row_group_per_episode(table, path)
-    else:
-        # No episode boundaries to align row groups to — keep a single write.
-        pq.write_table(table, str(path))
-
-
-def to_parquet_one_row_group_per_episode(df: pandas.DataFrame, path: Path) -> None:
-    """Write a (non-image) DataFrame to parquet with one row group per episode."""
-    table = pa.Table.from_pandas(df, preserve_index=False)
-    if "episode_index" in table.column_names:
-        write_table_one_row_group_per_episode(table, path)
-    else:
-        pq.write_table(table, str(path))
+    ds.to_parquet(path)
 
 
 def item_to_torch(item: dict) -> dict:

src/lerobot/robots/bi_openarm_follower/bi_openarm_follower.py

@@ -18,8 +18,7 @@ import logging
 from functools import cached_property
 
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.bimanual import BimanualMixin
-from lerobot.utils.decorators import check_if_not_connected
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 
 from ..openarm_follower import OpenArmFollower, OpenArmFollowerConfig
 from ..robot import Robot
@@ -28,7 +27,7 @@ from .config_bi_openarm_follower import BiOpenArmFollowerConfig
 logger = logging.getLogger(__name__)
 
 
-class BiOpenArmFollower(BimanualMixin, Robot):
+class BiOpenArmFollower(Robot):
     """
     Bimanual OpenArm Follower Arms
     """
@@ -40,17 +39,15 @@ class BiOpenArmFollower(BimanualMixin, Robot):
         super().__init__(config)
         self.config = config
 
-        # Top-level cameras are opened by `left_arm` for convenience, but their
-        # keys stay unprefixed in observations (tracked via `_top_level_cam_keys`).
-        self._top_level_cam_keys = set(config.cameras)
-        _collisions = self._top_level_cam_keys & set(
-            config.left_arm_config.cameras
-        ) | self._top_level_cam_keys & set(config.right_arm_config.cameras)
-        if _collisions:
-            raise ValueError(
-                f"Top-level camera names collide with per-arm camera names: {sorted(_collisions)}"
-            )
-        left_arm_cameras = {**config.left_arm_config.cameras, **config.cameras}
+        # Top-level cameras are distributed evenly: each arm's OpenArmFollower
+        # will only open the cameras assigned to it. Per-arm cameras are used
+        # as fallback when top-level cameras are empty.
+        if config.cameras:
+            left_cameras = config.cameras
+            right_cameras = {}
+        else:
+            left_cameras = config.left_arm_config.cameras
+            right_cameras = config.right_arm_config.cameras
 
         left_arm_config = OpenArmFollowerConfig(
             id=f"{config.id}_left" if config.id else None,
@@ -59,7 +56,7 @@ class BiOpenArmFollower(BimanualMixin, Robot):
             disable_torque_on_disconnect=config.left_arm_config.disable_torque_on_disconnect,
             use_velocity_and_torque=config.left_arm_config.use_velocity_and_torque,
             max_relative_target=config.left_arm_config.max_relative_target,
-            cameras=left_arm_cameras,
+            cameras=left_cameras,
             side=config.left_arm_config.side,
             can_interface=config.left_arm_config.can_interface,
             use_can_fd=config.left_arm_config.use_can_fd,
@@ -78,7 +75,7 @@ class BiOpenArmFollower(BimanualMixin, Robot):
             disable_torque_on_disconnect=config.right_arm_config.disable_torque_on_disconnect,
             use_velocity_and_torque=config.right_arm_config.use_velocity_and_torque,
             max_relative_target=config.right_arm_config.max_relative_target,
-            cameras=config.right_arm_config.cameras,
+            cameras=right_cameras,
             side=config.right_arm_config.side,
             can_interface=config.right_arm_config.can_interface,
             use_can_fd=config.right_arm_config.use_can_fd,
@@ -98,19 +95,22 @@ class BiOpenArmFollower(BimanualMixin, Robot):
 
     @property
     def _motors_ft(self) -> dict[str, type]:
+        left_arm_motors_ft = self.left_arm._motors_ft
+        right_arm_motors_ft = self.right_arm._motors_ft
+
+        # Right first, then left — matches the teleoperator (OpenArmMini) ordering
+        # and the dataset feature names recorded during data collection.
         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()},
+            **{f"right_{k}": v for k, v in right_arm_motors_ft.items()},
+            **{f"left_{k}": v for k, v in left_arm_motors_ft.items()},
         }
 
     @property
     def _cameras_ft(self) -> dict[str, tuple]:
-        out: dict[str, tuple] = {}
-        for k, v in self.left_arm._cameras_ft.items():
-            out[k if k in self._top_level_cam_keys else f"left_{k}"] = v
-        for k, v in self.right_arm._cameras_ft.items():
-            out[f"right_{k}"] = v
-        return out
+        # Cameras already have unique user-chosen names (e.g. "left_wrist", "base",
+        # "right_wrist"), so we merge them directly — unlike motors which need the
+        # left_/right_ prefix to disambiguate identical per-arm joint names.
+        return {**self.left_arm._cameras_ft, **self.right_arm._cameras_ft}
 
     @cached_property
     def observation_features(self) -> dict[str, type | tuple]:
@@ -120,6 +120,27 @@ class BiOpenArmFollower(BimanualMixin, Robot):
     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()
+
     def setup_motors(self) -> None:
         raise NotImplementedError(
             "Motor ID configuration is typically done via manufacturer tools for CAN motors."
@@ -127,15 +148,21 @@ class BiOpenArmFollower(BimanualMixin, Robot):
 
     @check_if_not_connected
     def get_observation(self) -> RobotObservation:
-        obs_dict: RobotObservation = {}
+        obs_dict = {}
 
-        # Add "left_" prefix to per-arm keys; keep top-level camera keys unprefixed.
-        for key, value in self.left_arm.get_observation().items():
-            obs_dict[key if key in self._top_level_cam_keys else f"left_{key}"] = value
+        # Camera keys that should NOT get the arm prefix (they already have unique names)
+        left_cam_keys = set(self.left_arm.cameras.keys())
+        right_cam_keys = set(self.right_arm.cameras.keys())
 
-        # Add "right_" prefix
-        for key, value in self.right_arm.get_observation().items():
-            obs_dict[f"right_{key}"] = value
+        # Right first, then left — matches the teleoperator (OpenArmMini) ordering
+        # and the dataset feature names recorded during data collection.
+        right_obs = self.right_arm.get_observation()
+        for key, value in right_obs.items():
+            obs_dict[key if key in right_cam_keys else f"right_{key}"] = value
+
+        left_obs = self.left_arm.get_observation()
+        for key, value in left_obs.items():
+            obs_dict[key if key in left_cam_keys else f"left_{key}"] = value
 
         return obs_dict
 
@@ -162,4 +189,9 @@ class BiOpenArmFollower(BimanualMixin, Robot):
         prefixed_sent_action_left = {f"left_{key}": value for key, value in sent_action_left.items()}
         prefixed_sent_action_right = {f"right_{key}": value for key, value in sent_action_right.items()}
 
-        return {**prefixed_sent_action_left, **prefixed_sent_action_right}
+        return {**prefixed_sent_action_right, **prefixed_sent_action_left}
+
+    @check_if_not_connected
+    def disconnect(self):
+        self.left_arm.disconnect()
+        self.right_arm.disconnect()

src/lerobot/robots/bi_openarm_follower/config_bi_openarm_follower.py

@@ -32,7 +32,5 @@ class BiOpenArmFollowerConfig(RobotConfig):
     left_arm_config: OpenArmFollowerConfigBase
     right_arm_config: OpenArmFollowerConfigBase
 
-    # Top-level cameras not attached to a specific side. Keys are kept as-is in
-    # observations (no `left_`/`right_` prefix). Per-arm cameras (declared on
-    # `{left,right}_arm_config.cameras`) are prefixed.
+    # Top-level cameras shared across both arms.
     cameras: dict[str, CameraConfig] = field(default_factory=dict)

src/lerobot/robots/bi_rebot_b601_follower/bi_rebot_b601_follower.py

@@ -18,8 +18,7 @@ import logging
 from functools import cached_property
 
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.bimanual import BimanualMixin
-from lerobot.utils.decorators import check_if_not_connected
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 
 from ..rebot_b601_follower import RebotB601Follower, RebotB601FollowerRobotConfig
 from ..robot import Robot
@@ -28,7 +27,7 @@ from .config_bi_rebot_b601_follower import BiRebotB601FollowerConfig
 logger = logging.getLogger(__name__)
 
 
-class BiRebotB601Follower(BimanualMixin, Robot):
+class BiRebotB601Follower(Robot):
     """Bimanual Seeed Studio reBot B601-DM follower.
 
     Composes two single-arm :class:`RebotB601Follower` instances. Observation and
@@ -42,18 +41,6 @@ class BiRebotB601Follower(BimanualMixin, Robot):
         super().__init__(config)
         self.config = config
 
-        # Top-level cameras are opened by `left_arm` for convenience, but their
-        # keys stay unprefixed in observations (tracked via `_top_level_cam_keys`).
-        self._top_level_cam_keys = set(config.cameras)
-        _collisions = self._top_level_cam_keys & set(
-            config.left_arm_config.cameras
-        ) | self._top_level_cam_keys & set(config.right_arm_config.cameras)
-        if _collisions:
-            raise ValueError(
-                f"Top-level camera names collide with per-arm camera names: {sorted(_collisions)}"
-            )
-        left_arm_cameras = {**config.left_arm_config.cameras, **config.cameras}
-
         left_arm_config = RebotB601FollowerRobotConfig(
             id=f"{config.id}_left" if config.id else None,
             calibration_dir=config.calibration_dir,
@@ -62,7 +49,7 @@ class BiRebotB601Follower(BimanualMixin, Robot):
             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=left_arm_cameras,
+            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,
@@ -99,12 +86,10 @@ class BiRebotB601Follower(BimanualMixin, Robot):
 
     @property
     def _cameras_ft(self) -> dict[str, tuple]:
-        out: dict[str, tuple] = {}
-        for k, v in self.left_arm._cameras_ft.items():
-            out[k if k in self._top_level_cam_keys else f"left_{k}"] = v
-        for k, v in self.right_arm._cameras_ft.items():
-            out[f"right_{k}"] = v
-        return out
+        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]:
@@ -114,13 +99,32 @@ class BiRebotB601Follower(BimanualMixin, Robot):
     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: RobotObservation = {}
-        for k, v in self.left_arm.get_observation().items():
-            obs_dict[k if k in self._top_level_cam_keys else f"left_{k}"] = v
-        for k, v in self.right_arm.get_observation().items():
-            obs_dict[f"right_{k}"] = v
+        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
@@ -139,3 +143,8 @@ class BiRebotB601Follower(BimanualMixin, Robot):
             **{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

@@ -14,9 +14,7 @@
 # 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 dataclasses import dataclass
 
 from ..config import RobotConfig
 from ..rebot_b601_follower import RebotB601FollowerConfig
@@ -29,8 +27,3 @@ class BiRebotB601FollowerConfig(RobotConfig):
 
     left_arm_config: RebotB601FollowerConfig
     right_arm_config: RebotB601FollowerConfig
-
-    # Top-level cameras not attached to a specific side. Keys are kept as-is in
-    # observations (no `left_`/`right_` prefix). Per-arm cameras (declared on
-    # `{left,right}_arm_config.cameras`) are prefixed.
-    cameras: dict[str, CameraConfig] = field(default_factory=dict)

src/lerobot/robots/bi_so_follower/bi_so_follower.py

@@ -18,8 +18,7 @@ import logging
 from functools import cached_property
 
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.bimanual import BimanualMixin
-from lerobot.utils.decorators import check_if_not_connected
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 
 from ..robot import Robot
 from ..so_follower import SOFollower, SOFollowerRobotConfig
@@ -28,7 +27,7 @@ from .config_bi_so_follower import BiSOFollowerConfig
 logger = logging.getLogger(__name__)
 
 
-class BiSOFollower(BimanualMixin, Robot):
+class BiSOFollower(Robot):
     """
     [Bimanual SO Follower Arms](https://github.com/TheRobotStudio/SO-ARM100) designed by TheRobotStudio
     """
@@ -40,18 +39,6 @@ class BiSOFollower(BimanualMixin, Robot):
         super().__init__(config)
         self.config = config
 
-        # Top-level cameras are opened by `left_arm` for convenience, but their
-        # keys stay unprefixed in observations (tracked via `_top_level_cam_keys`).
-        self._top_level_cam_keys = set(config.cameras)
-        _collisions = self._top_level_cam_keys & set(
-            config.left_arm_config.cameras
-        ) | self._top_level_cam_keys & set(config.right_arm_config.cameras)
-        if _collisions:
-            raise ValueError(
-                f"Top-level camera names collide with per-arm camera names: {sorted(_collisions)}"
-            )
-        left_arm_cameras = {**config.left_arm_config.cameras, **config.cameras}
-
         left_arm_config = SOFollowerRobotConfig(
             id=f"{config.id}_left" if config.id else None,
             calibration_dir=config.calibration_dir,
@@ -59,7 +46,7 @@ class BiSOFollower(BimanualMixin, Robot):
             disable_torque_on_disconnect=config.left_arm_config.disable_torque_on_disconnect,
             max_relative_target=config.left_arm_config.max_relative_target,
             use_degrees=config.left_arm_config.use_degrees,
-            cameras=left_arm_cameras,
+            cameras=config.left_arm_config.cameras,
         )
 
         right_arm_config = SOFollowerRobotConfig(
@@ -90,12 +77,13 @@ class BiSOFollower(BimanualMixin, Robot):
 
     @property
     def _cameras_ft(self) -> dict[str, tuple]:
-        out: dict[str, tuple] = {}
-        for k, v in self.left_arm._cameras_ft.items():
-            out[k if k in self._top_level_cam_keys else f"left_{k}"] = v
-        for k, v in self.right_arm._cameras_ft.items():
-            out[f"right_{k}"] = v
-        return out
+        left_arm_cameras_ft = self.left_arm._cameras_ft
+        right_arm_cameras_ft = self.right_arm._cameras_ft
+
+        return {
+            **{f"left_{k}": v for k, v in left_arm_cameras_ft.items()},
+            **{f"right_{k}": v for k, v in right_arm_cameras_ft.items()},
+        }
 
     @cached_property
     def observation_features(self) -> dict[str, type | tuple]:
@@ -105,21 +93,42 @@ class BiSOFollower(BimanualMixin, Robot):
     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()
+
     def setup_motors(self) -> None:
         self.left_arm.setup_motors()
         self.right_arm.setup_motors()
 
     @check_if_not_connected
     def get_observation(self) -> RobotObservation:
-        obs_dict: RobotObservation = {}
+        obs_dict = {}
 
-        # Add "left_" prefix to per-arm keys; keep top-level camera keys unprefixed.
-        for key, value in self.left_arm.get_observation().items():
-            obs_dict[key if key in self._top_level_cam_keys else f"left_{key}"] = value
+        # Add "left_" prefix
+        left_obs = self.left_arm.get_observation()
+        obs_dict.update({f"left_{key}": value for key, value in left_obs.items()})
 
         # Add "right_" prefix
-        for key, value in self.right_arm.get_observation().items():
-            obs_dict[f"right_{key}"] = value
+        right_obs = self.right_arm.get_observation()
+        obs_dict.update({f"right_{key}": value for key, value in right_obs.items()})
 
         return obs_dict
 
@@ -142,3 +151,8 @@ class BiSOFollower(BimanualMixin, Robot):
         prefixed_sent_action_right = {f"right_{key}": value for key, value in sent_action_right.items()}
 
         return {**prefixed_sent_action_left, **prefixed_sent_action_right}
+
+    @check_if_not_connected
+    def disconnect(self):
+        self.left_arm.disconnect()
+        self.right_arm.disconnect()

src/lerobot/robots/bi_so_follower/config_bi_so_follower.py

@@ -14,9 +14,7 @@
 # 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 dataclasses import dataclass
 
 from ..config import RobotConfig
 from ..so_follower import SOFollowerConfig
@@ -29,8 +27,3 @@ class BiSOFollowerConfig(RobotConfig):
 
     left_arm_config: SOFollowerConfig
     right_arm_config: SOFollowerConfig
-
-    # Top-level cameras not attached to a specific side. Keys are kept as-is in
-    # observations (no `left_`/`right_` prefix). Per-arm cameras (declared on
-    # `{left,right}_arm_config.cameras`) are prefixed.
-    cameras: dict[str, CameraConfig] = field(default_factory=dict)

src/lerobot/scripts/lerobot_calibrate.py

@@ -54,7 +54,6 @@ from lerobot.teleoperators import (  # noqa: F401
     Teleoperator,
     TeleoperatorConfig,
     bi_openarm_leader,
-    bi_openarm_mini,
     bi_rebot_102_leader,
     bi_so_leader,
     homunculus,

src/lerobot/scripts/lerobot_find_joint_limits.py

@@ -57,7 +57,6 @@ from lerobot.robots import (  # noqa: F401
 from lerobot.teleoperators import (  # noqa: F401
     TeleoperatorConfig,
     bi_openarm_leader,
-    bi_openarm_mini,
     bi_rebot_102_leader,
     bi_so_leader,
     gamepad,

src/lerobot/scripts/lerobot_record.py

@@ -137,7 +137,6 @@ from lerobot.teleoperators import (  # noqa: F401
     Teleoperator,
     TeleoperatorConfig,
     bi_openarm_leader,
-    bi_openarm_mini,
     bi_rebot_102_leader,
     bi_so_leader,
     homunculus,

src/lerobot/scripts/lerobot_rollout.py

@@ -174,7 +174,6 @@ from lerobot.teleoperators import (  # noqa: F401
     Teleoperator,
     TeleoperatorConfig,
     bi_openarm_leader,
-    bi_openarm_mini,
     bi_rebot_102_leader,
     bi_so_leader,
     homunculus,

src/lerobot/scripts/lerobot_setup_motors.py

@@ -41,7 +41,6 @@ from lerobot.robots import (  # noqa: F401
 )
 from lerobot.teleoperators import (  # noqa: F401
     TeleoperatorConfig,
-    bi_openarm_mini,
     bi_rebot_102_leader,
     bi_so_leader,
     koch_leader,

src/lerobot/scripts/lerobot_teleoperate.py

@@ -89,7 +89,6 @@ from lerobot.teleoperators import (  # noqa: F401
     Teleoperator,
     TeleoperatorConfig,
     bi_openarm_leader,
-    bi_openarm_mini,
     bi_rebot_102_leader,
     bi_so_leader,
     gamepad,

src/lerobot/scripts/lerobot_train.py

@@ -45,7 +45,8 @@ from lerobot.common.train_utils import (
 from lerobot.common.wandb_utils import WandBLogger
 from lerobot.configs import parser
 from lerobot.configs.train import TrainPipelineConfig
-from lerobot.datasets import EpisodeAwareSampler, compute_sampler_state, make_dataset
+from lerobot.datasets import EpisodeAwareSampler, compute_sampler_state
+from lerobot.datasets.factory import make_train_eval_datasets
 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
@@ -244,19 +245,19 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     # LeRobotDataset skips its snapshot_download when try_load() succeeds, so no rank re-downloads.
     if is_main_process:
         logging.info("Creating dataset")
-        dataset = make_dataset(cfg)
+        dataset, eval_dataset = make_train_eval_datasets(cfg)
 
     accelerator.wait_for_everyone()
 
     # Other ranks read from the shared copy populated by the main process.
     if not is_main_process:
-        dataset = make_dataset(cfg)
+        dataset, eval_dataset = make_train_eval_datasets(cfg)
 
     # Create environment used for evaluating checkpoints during training on simulation data.
     # On real-world data, no need to create an environment as evaluations are done outside train.py,
     # using the eval.py instead, with gym_dora environment and dora-rs.
     eval_env = None
-    if cfg.eval_freq > 0 and cfg.env is not None and is_main_process:
+    if cfg.env_eval_freq > 0 and cfg.env is not None and is_main_process:
         logging.info("Creating env")
         eval_env = make_env(cfg.env, n_envs=cfg.eval.batch_size, use_async_envs=cfg.eval.use_async_envs)
 
@@ -434,6 +435,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
                     f"Resuming data order at epoch {sampler_state['epoch']}, "
                     f"sample {sampler_state['start_index']}"
                 )
+        if dataset.reader._absolute_to_relative_idx is not None:
+            sampler.indices = [dataset.reader._absolute_to_relative_idx[i] for i in sampler.indices]
     else:
         shuffle = True
         sampler = None
@@ -455,6 +458,31 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         persistent_workers=cfg.persistent_workers and cfg.num_workers > 0,
     )
 
+    # Build eval dataloader if a held-out split exists
+    eval_dataloader = None
+    if eval_dataset is not None:
+        eval_ds = eval_dataset
+        if cfg.max_eval_samples > 0 and hasattr(eval_dataset, "hf_dataset"):
+            task_indices = eval_dataset.hf_dataset["task_index"]
+            unique_tasks = sorted(set(task_indices))
+            per_task = max(1, cfg.max_eval_samples // len(unique_tasks))
+            selected: list[int] = []
+            for t in unique_tasks:
+                frames = [i for i, ti in enumerate(task_indices) if ti == t][:per_task]
+                selected.extend(frames)
+            eval_ds = torch.utils.data.Subset(eval_dataset, selected)
+
+        eval_collate_fn = lerobot_collate_fn if dataset.meta.has_language_columns else None
+        eval_dataloader = torch.utils.data.DataLoader(
+            eval_ds,
+            batch_size=cfg.batch_size,
+            shuffle=False,
+            num_workers=cfg.num_workers,
+            pin_memory=device.type == "cuda",
+            drop_last=False,
+            collate_fn=eval_collate_fn,
+        )
+
     # Prepare everything with accelerator
     accelerator.wait_for_everyone()
     policy, optimizer, dataloader, lr_scheduler = accelerator.prepare(
@@ -534,7 +562,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         train_tracker.step()
         is_log_step = cfg.log_freq > 0 and step % cfg.log_freq == 0
         is_saving_step = step % cfg.save_freq == 0 or step == cfg.steps
-        is_eval_step = cfg.eval_freq > 0 and step % cfg.eval_freq == 0
+        is_env_eval_step = cfg.env_eval_freq > 0 and step % cfg.env_eval_freq == 0
+        is_eval_step = cfg.eval_steps > 0 and eval_dataloader is not None and step % cfg.eval_steps == 0
 
         if is_log_step:
             # Collective reduce must run on every rank, before the main-process gate below.
@@ -557,6 +586,27 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
                     wandb_logger.log_dict(wandb_log_dict, step)
             train_tracker.reset_averages()
 
+        if is_eval_step:
+            policy.eval()
+            eval_loss_sum = 0.0
+            n_eval_batches = 0
+            with torch.no_grad(), accelerator.autocast():
+                for eval_batch in eval_dataloader:
+                    for cam_key in dataset.meta.camera_keys:
+                        if cam_key in eval_batch and eval_batch[cam_key].dtype == torch.uint8:
+                            eval_batch[cam_key] = eval_batch[cam_key].to(dtype=torch.float32) / 255.0
+                    eval_batch = preprocessor(eval_batch)
+                    loss, _ = policy.forward(eval_batch)
+                    eval_loss_sum += loss.item()
+                    n_eval_batches += 1
+            eval_loss = eval_loss_sum / max(n_eval_batches, 1)
+            policy.train()
+
+            if is_main_process:
+                logging.info(f"step {step}: eval_loss={eval_loss:.4f}")
+                if wandb_logger:
+                    wandb_logger.log_dict({"eval_loss": eval_loss}, step=step, mode="eval")
+
         if cfg.save_checkpoint and is_saving_step:
             if is_main_process:
                 logging.info(f"Checkpoint policy after step {step}")
@@ -579,7 +629,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
 
             accelerator.wait_for_everyone()
 
-        if cfg.env and is_eval_step:
+        if cfg.env and is_env_eval_step:
             if is_main_process:
                 step_id = get_step_identifier(step, cfg.steps)
                 logging.info(f"Eval policy at step {step}")

src/lerobot/teleoperators/bi_openarm_leader/bi_openarm_leader.py

@@ -18,8 +18,7 @@ import logging
 from functools import cached_property
 
 from lerobot.types import RobotAction
-from lerobot.utils.bimanual import BimanualMixin
-from lerobot.utils.decorators import check_if_not_connected
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 
 from ..openarm_leader import OpenArmLeader, OpenArmLeaderConfig
 from ..teleoperator import Teleoperator
@@ -28,7 +27,7 @@ from .config_bi_openarm_leader import BiOpenArmLeaderConfig
 logger = logging.getLogger(__name__)
 
 
-class BiOpenArmLeader(BimanualMixin, Teleoperator):
+class BiOpenArmLeader(Teleoperator):
     """
     Bimanual OpenArm Leader Arms
     """
@@ -87,6 +86,27 @@ class BiOpenArmLeader(BimanualMixin, Teleoperator):
     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()
+
     def setup_motors(self) -> None:
         raise NotImplementedError(
             "Motor ID configuration is typically done via manufacturer tools for CAN motors."
@@ -109,3 +129,8 @@ class BiOpenArmLeader(BimanualMixin, Teleoperator):
     def send_feedback(self, feedback: dict[str, float]) -> None:
         # TODO: Implement force feedback
         raise NotImplementedError
+
+    @check_if_not_connected
+    def disconnect(self) -> None:
+        self.left_arm.disconnect()
+        self.right_arm.disconnect()

src/lerobot/teleoperators/bi_openarm_leader/config_bi_openarm_leader.py

@@ -23,7 +23,7 @@ from ..openarm_leader import OpenArmLeaderConfigBase
 @TeleoperatorConfig.register_subclass("bi_openarm_leader")
 @dataclass
 class BiOpenArmLeaderConfig(TeleoperatorConfig):
-    """Configuration class for Bi OpenArm Leader teleoperators."""
+    """Configuration class for Bi OpenArm Follower robots."""
 
     left_arm_config: OpenArmLeaderConfigBase
     right_arm_config: OpenArmLeaderConfigBase

src/lerobot/teleoperators/bi_openarm_mini/__init__.py

@@ -1,20 +0,0 @@
-#!/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_openarm_mini import BiOpenArmMini
-from .config_bi_openarm_mini import BiOpenArmMiniConfig
-
-__all__ = ["BiOpenArmMini", "BiOpenArmMiniConfig"]

src/lerobot/teleoperators/bi_openarm_mini/bi_openarm_mini.py

@@ -1,101 +0,0 @@
-#!/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.bimanual import BimanualMixin
-from lerobot.utils.decorators import check_if_not_connected
-
-from ..openarm_mini import OpenArmMini, OpenArmMiniConfig
-from ..teleoperator import Teleoperator
-from .config_bi_openarm_mini import BiOpenArmMiniConfig
-
-logger = logging.getLogger(__name__)
-
-
-class BiOpenArmMini(BimanualMixin, Teleoperator):
-    """Bimanual OpenArm Mini teleoperator.
-
-    Composes two single-arm :class:`OpenArmMini` instances. Action and feedback
-    keys of each arm are namespaced with a ``left_`` / ``right_`` prefix, so a
-    bimanual leader can teleoperate a bimanual OpenArm follower.
-    """
-
-    config_class = BiOpenArmMiniConfig
-    name = "bi_openarm_mini"
-
-    def __init__(self, config: BiOpenArmMiniConfig):
-        super().__init__(config)
-        self.config = config
-
-        # `side` is forced to match left/right regardless of what the user passed
-        # on the per-arm base config — the bimanual wrapper owns the side semantics.
-        left_arm_config = OpenArmMiniConfig(
-            id=f"{config.id}_left" if config.id else None,
-            calibration_dir=config.calibration_dir,
-            port=config.left_arm_config.port,
-            side="left",
-            use_degrees=config.left_arm_config.use_degrees,
-        )
-
-        right_arm_config = OpenArmMiniConfig(
-            id=f"{config.id}_right" if config.id else None,
-            calibration_dir=config.calibration_dir,
-            port=config.right_arm_config.port,
-            side="right",
-            use_degrees=config.right_arm_config.use_degrees,
-        )
-
-        self.left_arm = OpenArmMini(left_arm_config)
-        self.right_arm = OpenArmMini(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 {
-            **{f"left_{k}": v for k, v in self.left_arm.feedback_features.items()},
-            **{f"right_{k}": v for k, v in self.right_arm.feedback_features.items()},
-        }
-
-    def setup_motors(self) -> None:
-        self.left_arm.setup_motors()
-        self.right_arm.setup_motors()
-
-    @check_if_not_connected
-    def get_action(self) -> RobotAction:
-        action: RobotAction = {}
-        for k, v in self.left_arm.get_action().items():
-            action[f"left_{k}"] = v
-        for k, v in self.right_arm.get_action().items():
-            action[f"right_{k}"] = v
-        return action
-
-    @check_if_not_connected
-    def send_feedback(self, feedback: dict[str, float]) -> None:
-        left_fb = {k.removeprefix("left_"): v for k, v in feedback.items() if k.startswith("left_")}
-        right_fb = {k.removeprefix("right_"): v for k, v in feedback.items() if k.startswith("right_")}
-        if left_fb:
-            self.left_arm.send_feedback(left_fb)
-        if right_fb:
-            self.right_arm.send_feedback(right_fb)

src/lerobot/teleoperators/bi_openarm_mini/config_bi_openarm_mini.py

@@ -1,29 +0,0 @@
-#!/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 ..openarm_mini import OpenArmMiniConfigBase
-
-
-@TeleoperatorConfig.register_subclass("bi_openarm_mini")
-@dataclass
-class BiOpenArmMiniConfig(TeleoperatorConfig):
-    """Configuration class for Bi OpenArm Mini teleoperators."""
-
-    left_arm_config: OpenArmMiniConfigBase
-    right_arm_config: OpenArmMiniConfigBase

src/lerobot/teleoperators/bi_rebot_102_leader/__init__.py

@@ -14,7 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .bi_rebot_102_leader import BiRebot102Leader
-from .config_bi_rebot_102_leader import BiRebot102LeaderConfig
+from .bi_rebot_102_leader import BiRebotArm102Leader
+from .config_bi_rebot_102_leader import BiRebotArm102LeaderConfig
 
-__all__ = ["BiRebot102Leader", "BiRebot102LeaderConfig"]
+__all__ = ["BiRebotArm102Leader", "BiRebotArm102LeaderConfig"]

src/lerobot/teleoperators/bi_rebot_102_leader/bi_rebot_102_leader.py

@@ -18,17 +18,16 @@ import logging
 from functools import cached_property
 
 from lerobot.types import RobotAction
-from lerobot.utils.bimanual import BimanualMixin
-from lerobot.utils.decorators import check_if_not_connected
+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 BiRebot102LeaderConfig
+from .config_bi_rebot_102_leader import BiRebotArm102LeaderConfig
 
 logger = logging.getLogger(__name__)
 
 
-class BiRebot102Leader(BimanualMixin, Teleoperator):
+class BiRebotArm102Leader(Teleoperator):
     """Bimanual Seeed Studio StarArm102 / reBot Arm 102 leader.
 
     Composes two single-arm :class:`RebotArm102Leader` instances. Action keys of
@@ -36,10 +35,10 @@ class BiRebot102Leader(BimanualMixin, Teleoperator):
     leader can teleoperate a bimanual reBot B601 follower.
     """
 
-    config_class = BiRebot102LeaderConfig
+    config_class = BiRebotArm102LeaderConfig
     name = "bi_rebot_102_leader"
 
-    def __init__(self, config: BiRebot102LeaderConfig):
+    def __init__(self, config: BiRebotArm102LeaderConfig):
         super().__init__(config)
         self.config = config
 
@@ -77,6 +76,27 @@ class BiRebot102Leader(BimanualMixin, Teleoperator):
     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 = {}
@@ -86,3 +106,8 @@ class BiRebot102Leader(BimanualMixin, Teleoperator):
 
     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

@@ -22,7 +22,7 @@ from ..rebot_102_leader import RebotArm102LeaderConfig
 
 @TeleoperatorConfig.register_subclass("bi_rebot_102_leader")
 @dataclass
-class BiRebot102LeaderConfig(TeleoperatorConfig):
+class BiRebotArm102LeaderConfig(TeleoperatorConfig):
     """Configuration class for the bimanual reBot Arm 102 leader teleoperator."""
 
     left_arm_config: RebotArm102LeaderConfig

src/lerobot/teleoperators/bi_so_leader/bi_so_leader.py

@@ -17,9 +17,7 @@
 import logging
 from functools import cached_property
 
-from lerobot.types import RobotAction
-from lerobot.utils.bimanual import BimanualMixin
-from lerobot.utils.decorators import check_if_not_connected
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 
 from ..so_leader import SOLeader, SOLeaderTeleopConfig
 from ..teleoperator import Teleoperator
@@ -28,7 +26,7 @@ from .config_bi_so_leader import BiSOLeaderConfig
 logger = logging.getLogger(__name__)
 
 
-class BiSOLeader(BimanualMixin, Teleoperator):
+class BiSOLeader(Teleoperator):
     """
     [Bimanual SO Leader Arms](https://github.com/TheRobotStudio/SO-ARM100) designed by TheRobotStudio
     """
@@ -69,12 +67,33 @@ class BiSOLeader(BimanualMixin, Teleoperator):
     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()
+
     def setup_motors(self) -> None:
         self.left_arm.setup_motors()
         self.right_arm.setup_motors()
 
     @check_if_not_connected
-    def get_action(self) -> RobotAction:
+    def get_action(self) -> dict[str, float]:
         action_dict = {}
 
         # Add "left_" prefix
@@ -90,3 +109,8 @@ class BiSOLeader(BimanualMixin, Teleoperator):
     def send_feedback(self, feedback: dict[str, float]) -> None:
         # TODO: Implement force feedback
         raise NotImplementedError
+
+    @check_if_not_connected
+    def disconnect(self) -> None:
+        self.left_arm.disconnect()
+        self.right_arm.disconnect()

src/lerobot/teleoperators/openarm_mini/__init__.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python
 
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+# Copyright 2025 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.
@@ -14,7 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .config_openarm_mini import OpenArmMiniConfig, OpenArmMiniConfigBase
+from .config_openarm_mini import OpenArmMiniConfig
 from .openarm_mini import OpenArmMini
 
-__all__ = ["OpenArmMini", "OpenArmMiniConfig", "OpenArmMiniConfigBase"]
+__all__ = ["OpenArmMini", "OpenArmMiniConfig"]

src/lerobot/teleoperators/openarm_mini/config_openarm_mini.py

@@ -19,21 +19,12 @@ from dataclasses import dataclass
 from ..config import TeleoperatorConfig
 
 
-@dataclass
-class OpenArmMiniConfigBase:
-    """Base configuration for the OpenArm Mini teleoperator (Feetech STS3215, 7DOF + gripper)."""
-
-    # Serial port for the Feetech bus (e.g., "/dev/ttyUSB0").
-    port: str
-
-    # Side of the arm: "left" or "right". Controls per-joint direction flips applied
-    # during readout. If `None`, no flipping is applied.
-    side: str | None = None
-
-    use_degrees: bool = True
-
-
 @TeleoperatorConfig.register_subclass("openarm_mini")
 @dataclass
-class OpenArmMiniConfig(TeleoperatorConfig, OpenArmMiniConfigBase):
-    pass
+class OpenArmMiniConfig(TeleoperatorConfig):
+    """Configuration for OpenArm Mini teleoperator with Feetech motors (dual arms)."""
+
+    port_right: str = "/dev/ttyUSB0"
+    port_left: str = "/dev/ttyUSB1"
+
+    use_degrees: bool = True

src/lerobot/teleoperators/openarm_mini/openarm_mini.py

@@ -31,22 +31,22 @@ from .config_openarm_mini import OpenArmMiniConfig
 
 logger = logging.getLogger(__name__)
 
-# Per-side motor direction flips applied during readout.
-SIDE_MOTORS_TO_FLIP: dict[str, list[str]] = {
-    "left": ["joint_1", "joint_3", "joint_4", "joint_5", "joint_6", "joint_7"],
-    "right": ["joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_7"],
-}
+# Motors whose direction is inverted during readout
+RIGHT_MOTORS_TO_FLIP = ["joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_7"]
+LEFT_MOTORS_TO_FLIP = ["joint_1", "joint_3", "joint_4", "joint_5", "joint_6", "joint_7"]
 
-# Leader joint 6 ↔ follower joint 7 (symmetric — its own inverse).
+# Leader joint 6 maps to follower joint 7 and vice versa
 JOINT_REMAP = {"joint_6": "joint_7", "joint_7": "joint_6"}
+JOINT_REMAP_REVERSE = {"joint_7": "joint_6", "joint_6": "joint_7"}
 
 GRIPPER_TELEOP_TO_DEGREES = -0.65
 
 
 class OpenArmMini(Teleoperator):
-    """OpenArm Mini single-arm teleoperator (Feetech STS3215, 7DOF + gripper).
+    """
+    OpenArm Mini Teleoperator with dual Feetech-based arms (8 motors per arm).
 
-    For the bimanual setup, see :class:`BiOpenArmMini` which composes two of these.
+    Each arm has 7 joints plus a gripper, using Feetech STS3215 servos.
     """
 
     config_class = OpenArmMiniConfig
@@ -56,12 +56,9 @@ class OpenArmMini(Teleoperator):
         super().__init__(config)
         self.config = config
 
-        if config.side is not None and config.side not in SIDE_MOTORS_TO_FLIP:
-            raise ValueError(f"Invalid side '{config.side}'; expected 'left', 'right', or None.")
-        self._motors_to_flip: list[str] = SIDE_MOTORS_TO_FLIP.get(config.side, []) if config.side else []
-
         norm_mode_body = MotorNormMode.DEGREES
-        motors = {
+
+        motors_right = {
             "joint_1": Motor(1, "sts3215", norm_mode_body),
             "joint_2": Motor(2, "sts3215", norm_mode_body),
             "joint_3": Motor(3, "sts3215", norm_mode_body),
@@ -72,15 +69,46 @@ class OpenArmMini(Teleoperator):
             "gripper": Motor(8, "sts3215", MotorNormMode.RANGE_0_100),
         }
 
-        self.bus = FeetechMotorsBus(
-            port=self.config.port,
-            motors=motors,
-            calibration=self.calibration,
+        motors_left = {
+            "joint_1": Motor(1, "sts3215", norm_mode_body),
+            "joint_2": Motor(2, "sts3215", norm_mode_body),
+            "joint_3": Motor(3, "sts3215", norm_mode_body),
+            "joint_4": Motor(4, "sts3215", norm_mode_body),
+            "joint_5": Motor(5, "sts3215", norm_mode_body),
+            "joint_6": Motor(6, "sts3215", norm_mode_body),
+            "joint_7": Motor(7, "sts3215", norm_mode_body),
+            "gripper": Motor(8, "sts3215", MotorNormMode.RANGE_0_100),
+        }
+
+        cal_right = {
+            k.replace("right_", ""): v for k, v in (self.calibration or {}).items() if k.startswith("right_")
+        }
+        cal_left = {
+            k.replace("left_", ""): v for k, v in (self.calibration or {}).items() if k.startswith("left_")
+        }
+
+        self.bus_right = FeetechMotorsBus(
+            port=self.config.port_right,
+            motors=motors_right,
+            calibration=cal_right,
+        )
+
+        self.bus_left = FeetechMotorsBus(
+            port=self.config.port_left,
+            motors=motors_left,
+            calibration=cal_left,
         )
 
     @property
     def action_features(self) -> dict[str, type]:
-        return {f"{motor}.pos": float for motor in self.bus.motors}
+        # Right first, then left — matches the robot (BiOpenArmFollower) ordering
+        # and the dataset feature names recorded during data collection.
+        features: dict[str, type] = {}
+        for motor in self.bus_right.motors:
+            features[f"right_{motor}.pos"] = float
+        for motor in self.bus_left.motors:
+            features[f"left_{motor}.pos"] = float
+        return features
 
     @property
     def feedback_features(self) -> dict[str, type]:
@@ -88,12 +116,14 @@ class OpenArmMini(Teleoperator):
 
     @property
     def is_connected(self) -> bool:
-        return self.bus.is_connected
+        return self.bus_right.is_connected and self.bus_left.is_connected
 
     @check_if_already_connected
     def connect(self, calibrate: bool = True) -> None:
-        logger.info(f"Connecting arm on {self.config.port}...")
-        self.bus.connect()
+        logger.info(f"Connecting right arm on {self.config.port_right}...")
+        self.bus_right.connect()
+        logger.info(f"Connecting left arm on {self.config.port_left}...")
+        self.bus_left.connect()
 
         if calibrate:
             self.calibrate()
@@ -103,14 +133,14 @@ class OpenArmMini(Teleoperator):
 
     @property
     def is_calibrated(self) -> bool:
-        return self.bus.is_calibrated
+        return self.bus_right.is_calibrated and self.bus_left.is_calibrated
 
     def calibrate(self) -> None:
         """
-        Run calibration procedure for a single OpenArm Mini arm.
+        Run calibration procedure for OpenArm Mini.
 
         1. Disable torque
-        2. Ask user to position arm in hanging position with gripper closed
+        2. Ask user to position arms in hanging position with grippers closed
         3. Set this as zero position via half-turn homing
         4. Interactive gripper calibration (open/close positions)
         5. Save calibration
@@ -122,51 +152,70 @@ class OpenArmMini(Teleoperator):
             )
             if user_input.strip().lower() != "c":
                 logger.info(f"Using existing calibration for {self.id}")
-                self.bus.write_calibration(self.calibration)
+                cal_right = {
+                    k.replace("right_", ""): v for k, v in self.calibration.items() if k.startswith("right_")
+                }
+                cal_left = {
+                    k.replace("left_", ""): v for k, v in self.calibration.items() if k.startswith("left_")
+                }
+                self.bus_right.write_calibration(cal_right)
+                self.bus_left.write_calibration(cal_left)
                 return
 
         logger.info(f"\nRunning calibration for {self}")
 
-        self.bus.disable_torque()
+        self._calibrate_arm("right", self.bus_right)
+        self._calibrate_arm("left", self.bus_left)
 
-        logger.info("Setting Phase to 12 for all motors...")
-        for motor in self.bus.motors:
-            self.bus.write("Phase", motor, 12)
+        self._save_calibration()
+        print(f"\nCalibration complete and saved to {self.calibration_fpath}")
 
-        for motor in self.bus.motors:
-            self.bus.write("Operating_Mode", motor, OperatingMode.POSITION.value)
+    def _calibrate_arm(self, arm_name: str, bus: FeetechMotorsBus) -> None:
+        """Calibrate a single arm with Feetech motors."""
+        logger.info(f"\n=== Calibrating {arm_name.upper()} arm ===")
+
+        bus.disable_torque()
+
+        logger.info(f"Setting Phase to 12 for all motors in {arm_name.upper()} arm...")
+        for motor in bus.motors:
+            bus.write("Phase", motor, 12)
+
+        for motor in bus.motors:
+            bus.write("Operating_Mode", motor, OperatingMode.POSITION.value)
 
         input(
-            "\nCalibration: Zero Position\n"
+            f"\nCalibration: Zero Position ({arm_name.upper()} arm)\n"
             "Position the arm in the following configuration:\n"
             "  - Arm hanging straight down\n"
             "  - Gripper closed\n"
             "Press ENTER when ready..."
         )
 
-        homing_offsets = self.bus.set_half_turn_homings()
-        logger.info("Arm zero position set.")
+        homing_offsets = bus.set_half_turn_homings()
+        logger.info(f"{arm_name.capitalize()} arm zero position set.")
 
-        print("\nSetting motor ranges\n")
+        print(f"\nSetting motor ranges for {arm_name.upper()} arm\n")
 
         if self.calibration is None:
             self.calibration = {}
 
-        motor_resolution = self.bus.model_resolution_table[list(self.bus.motors.values())[0].model]
+        motor_resolution = bus.model_resolution_table[list(bus.motors.values())[0].model]
         max_res = motor_resolution - 1
 
-        for motor_name, motor in self.bus.motors.items():
+        for motor_name, motor in bus.motors.items():
+            prefixed_name = f"{arm_name}_{motor_name}"
+
             if motor_name == "gripper":
                 input(
-                    "\nGripper Calibration\n"
-                    "Step 1: CLOSE the gripper fully\n"
-                    "Press ENTER when gripper is closed..."
+                    f"\nGripper Calibration ({arm_name.upper()} arm)\n"
+                    f"Step 1: CLOSE the gripper fully\n"
+                    f"Press ENTER when gripper is closed..."
                 )
-                closed_pos = self.bus.read("Present_Position", motor_name, normalize=False)
+                closed_pos = bus.read("Present_Position", motor_name, normalize=False)
                 logger.info(f"  Gripper closed position recorded: {closed_pos}")
 
                 input("\nStep 2: OPEN the gripper fully\nPress ENTER when gripper is fully open...")
-                open_pos = self.bus.read("Present_Position", motor_name, normalize=False)
+                open_pos = bus.read("Present_Position", motor_name, normalize=False)
                 logger.info(f"  Gripper open position recorded: {open_pos}")
 
                 if closed_pos < open_pos:
@@ -179,16 +228,16 @@ class OpenArmMini(Teleoperator):
                     drive_mode = 1
 
                 logger.info(
-                    f"  {motor_name}: range set to [{range_min}, {range_max}] "
+                    f"  {prefixed_name}: range set to [{range_min}, {range_max}] "
                     f"(0=closed, 100=open, drive_mode={drive_mode})"
                 )
             else:
                 range_min = 0
                 range_max = max_res
                 drive_mode = 0
-                logger.info(f"  {motor_name}: range set to [0, {max_res}] (full motor range)")
+                logger.info(f"  {prefixed_name}: range set to [0, {max_res}] (full motor range)")
 
-            self.calibration[motor_name] = MotorCalibration(
+            self.calibration[prefixed_name] = MotorCalibration(
                 id=motor.id,
                 drive_mode=drive_mode,
                 homing_offset=homing_offsets[motor_name],
@@ -196,68 +245,108 @@ class OpenArmMini(Teleoperator):
                 range_max=range_max,
             )
 
-        self.bus.write_calibration(self.calibration)
-        self._save_calibration()
-        print(f"\nCalibration complete and saved to {self.calibration_fpath}")
+        cal_for_bus = {
+            k.replace(f"{arm_name}_", ""): v
+            for k, v in self.calibration.items()
+            if k.startswith(f"{arm_name}_")
+        }
+        bus.write_calibration(cal_for_bus)
 
     def configure(self) -> None:
-        self.bus.disable_torque()
-        self.bus.configure_motors()
-        for motor in self.bus.motors:
-            self.bus.write("Operating_Mode", motor, OperatingMode.POSITION.value)
+        self.bus_right.disable_torque()
+        self.bus_right.configure_motors()
+        for motor in self.bus_right.motors:
+            self.bus_right.write("Operating_Mode", motor, OperatingMode.POSITION.value)
+
+        self.bus_left.disable_torque()
+        self.bus_left.configure_motors()
+        for motor in self.bus_left.motors:
+            self.bus_left.write("Operating_Mode", motor, OperatingMode.POSITION.value)
 
     def setup_motors(self) -> None:
-        for motor in reversed(self.bus.motors):
-            input(f"Connect the controller board to the '{motor}' motor only and press enter.")
-            self.bus.setup_motor(motor)
-            print(f"'{motor}' motor id set to {self.bus.motors[motor].id}")
+        print("\nSetting up RIGHT arm motors...")
+        for motor in reversed(self.bus_right.motors):
+            input(f"Connect the controller board to the RIGHT '{motor}' motor only and press enter.")
+            self.bus_right.setup_motor(motor)
+            print(f"RIGHT '{motor}' motor id set to {self.bus_right.motors[motor].id}")
+
+        print("\nSetting up LEFT arm motors...")
+        for motor in reversed(self.bus_left.motors):
+            input(f"Connect the controller board to the LEFT '{motor}' motor only and press enter.")
+            self.bus_left.setup_motor(motor)
+            print(f"LEFT '{motor}' motor id set to {self.bus_left.motors[motor].id}")
 
     @check_if_not_connected
     def get_action(self) -> RobotAction:
-        """Get current action (read positions from all motors)."""
+        """Get current action from both arms (read positions from all motors)."""
         start = time.perf_counter()
 
-        positions = self.bus.sync_read("Present_Position")
+        right_positions = self.bus_right.sync_read("Present_Position")
+        left_positions = self.bus_left.sync_read("Present_Position")
 
+        # Right first, then left — matches the robot (BiOpenArmFollower) ordering
+        # and the dataset feature names recorded during data collection.
         # Joint 6↔7 remap: leader joint_6 → follower joint_7 and vice versa.
-        # Per-side direction flip is applied based on the configured `side`.
         action: dict[str, Any] = {}
-        for motor, val in positions.items():
+        for motor, val in right_positions.items():
             target = JOINT_REMAP.get(motor, motor)
             if motor == "gripper":
                 # Convert gripper from teleop 0-100 to openarms degrees: 0→0°, 100→-65°
-                action[f"{target}.pos"] = val * GRIPPER_TELEOP_TO_DEGREES
+                action[f"right_{target}.pos"] = val * GRIPPER_TELEOP_TO_DEGREES
             else:
-                action[f"{target}.pos"] = -val if motor in self._motors_to_flip else val
+                action[f"right_{target}.pos"] = -val if motor in RIGHT_MOTORS_TO_FLIP else val
+        for motor, val in left_positions.items():
+            target = JOINT_REMAP.get(motor, motor)
+            if motor == "gripper":
+                action[f"left_{target}.pos"] = val * GRIPPER_TELEOP_TO_DEGREES
+            else:
+                action[f"left_{target}.pos"] = -val if motor in LEFT_MOTORS_TO_FLIP else val
 
         dt_ms = (time.perf_counter() - start) * 1e3
         logger.debug(f"{self} read action: {dt_ms:.1f}ms")
         return action
 
     def enable_torque(self) -> None:
-        self.bus.enable_torque()
+        """Enable torque on both arms for position control."""
+        self.bus_right.enable_torque()
+        self.bus_left.enable_torque()
 
     def disable_torque(self) -> None:
-        self.bus.disable_torque()
+        """Disable torque on both arms for free movement."""
+        self.bus_right.disable_torque()
+        self.bus_left.disable_torque()
 
     def write_goal_positions(self, positions: dict[str, float]) -> None:
         """Write goal positions to motors (inverse of get_action flip/gripper/remap logic)."""
-        goals: dict[str, float] = {}
+        right_goals: dict[str, float] = {}
+        left_goals: dict[str, float] = {}
+
         for key, val in positions.items():
             if not key.endswith(".pos"):
                 continue
-            base = key.removesuffix(".pos")
-            # JOINT_REMAP is symmetric (its own inverse).
-            target = JOINT_REMAP.get(base, base)
-            if base == "gripper":
-                # Convert robot degrees to teleop 0-100: 0°→0, -65°→100
-                goals[target] = val / GRIPPER_TELEOP_TO_DEGREES
-            else:
-                # Un-flip using the ORIGINAL motor name (target = leader motor)
-                goals[target] = -val if target in self._motors_to_flip else val
+            motor_name = key.removesuffix(".pos")
+            if motor_name.startswith("right_"):
+                base = motor_name.removeprefix("right_")
+                # Reverse remap: follower joint_7 → leader joint_6 and vice versa
+                target = JOINT_REMAP_REVERSE.get(base, base)
+                if base == "gripper":
+                    # Convert robot degrees to teleop 0-100: 0°→0, -65°→100
+                    right_goals[target] = val / GRIPPER_TELEOP_TO_DEGREES
+                else:
+                    # Un-flip using the ORIGINAL motor name (target = leader motor)
+                    right_goals[target] = -val if target in RIGHT_MOTORS_TO_FLIP else val
+            elif motor_name.startswith("left_"):
+                base = motor_name.removeprefix("left_")
+                target = JOINT_REMAP_REVERSE.get(base, base)
+                if base == "gripper":
+                    left_goals[target] = val / GRIPPER_TELEOP_TO_DEGREES
+                else:
+                    left_goals[target] = -val if target in LEFT_MOTORS_TO_FLIP else val
 
-        if goals:
-            self.bus.sync_write("Goal_Position", goals)
+        if right_goals:
+            self.bus_right.sync_write("Goal_Position", right_goals)
+        if left_goals:
+            self.bus_left.sync_write("Goal_Position", left_goals)
 
     @check_if_not_connected
     def send_feedback(self, feedback: dict[str, float]) -> None:
@@ -265,5 +354,6 @@ class OpenArmMini(Teleoperator):
 
     @check_if_not_connected
     def disconnect(self) -> None:
-        self.bus.disconnect()
+        self.bus_right.disconnect()
+        self.bus_left.disconnect()
         logger.info(f"{self} disconnected.")

src/lerobot/teleoperators/utils.py

@@ -99,18 +99,14 @@ def make_teleoperator_from_config(config: TeleoperatorConfig) -> "Teleoperator":
         from .openarm_mini import OpenArmMini
 
         return OpenArmMini(config)
-    elif config.type == "bi_openarm_mini":
-        from .bi_openarm_mini import BiOpenArmMini
-
-        return BiOpenArmMini(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 BiRebot102Leader
+        from .bi_rebot_102_leader import BiRebotArm102Leader
 
-        return BiRebot102Leader(config)
+        return BiRebotArm102Leader(config)
     else:
         try:
             return cast("Teleoperator", make_device_from_device_class(config))

src/lerobot/utils/bimanual.py

@@ -1,63 +0,0 @@
-#!/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 typing import Any
-
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-
-
-class BimanualMixin:
-    """Lifecycle delegation for bimanual robots and teleoperators.
-
-    Concrete subclasses must populate ``self.left_arm`` and ``self.right_arm`` in
-    their own ``__init__``. They retain ownership of feature dicts and the
-    data-routing methods (``get_action`` / ``send_action`` / ``get_observation`` /
-    ``send_feedback``), which vary per-embodiment.
-
-    Inherit before the ``Robot`` / ``Teleoperator`` base so the mixin's methods
-    take precedence in the MRO::
-
-        class BiFooFollower(BimanualMixin, Robot): ...
-    """
-
-    left_arm: Any
-    right_arm: Any
-
-    @property
-    def is_connected(self) -> bool:
-        return self.left_arm.is_connected and self.right_arm.is_connected
-
-    @property
-    def is_calibrated(self) -> bool:
-        return self.left_arm.is_calibrated and self.right_arm.is_calibrated
-
-    @check_if_already_connected
-    def connect(self, calibrate: bool = True) -> None:
-        self.left_arm.connect(calibrate)
-        self.right_arm.connect(calibrate)
-
-    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 disconnect(self) -> None:
-        self.left_arm.disconnect()
-        self.right_arm.disconnect()

tests/annotations/test_writer.py

@@ -28,7 +28,6 @@ import pytest
 pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
 pytest.importorskip("pandas", reason="pandas is required (install lerobot[dataset])")
 
-import pandas as pd  # noqa: E402
 import pyarrow.parquet as pq  # noqa: E402
 
 from lerobot.annotations.steerable_pipeline.reader import iter_episodes  # noqa: E402
@@ -345,78 +344,6 @@ def test_annotation_metadata_sync_allows_non_streaming_load(
     assert len(dataset) == 24
 
 
-def _build_packed_dataset(root: Path, episode_lengths: list[int], *, fps: int = 10) -> Path:
-    """Pack several episodes into a single shard (vs build_annotation_dataset's one-per-file),
-    so the writer's rewrite must re-emit one row group per episode instead of collapsing them."""
-    from lerobot.datasets.io_utils import write_tasks
-    from lerobot.utils.io_utils import write_json
-
-    data_dir = root / "data" / "chunk-000"
-    data_dir.mkdir(parents=True, exist_ok=True)
-
-    episode_index, frame_index, timestamp, task_index, subtask_index = [], [], [], [], []
-    for ep, length in enumerate(episode_lengths):
-        episode_index += [ep] * length
-        frame_index += list(range(length))
-        timestamp += [round(i / fps, 6) for i in range(length)]
-        task_index += [0] * length
-        subtask_index += [0] * length  # legacy column the writer must drop
-    pd.DataFrame(
-        {
-            "episode_index": episode_index,
-            "frame_index": frame_index,
-            "timestamp": timestamp,
-            "task_index": task_index,
-            "subtask_index": subtask_index,
-        }
-    ).to_parquet(data_dir / "file-000.parquet", index=False)
-
-    tasks_df = pd.DataFrame({"task_index": [0]}, index=pd.Index(["do the thing"], name="task"))
-    write_tasks(tasks_df, root)
-    write_json(
-        {"codebase_version": "v3.1", "fps": fps, "features": {}, "total_episodes": len(episode_lengths)},
-        root / "meta" / "info.json",
-    )
-    return root
-
-
-def test_writer_one_row_group_per_episode(tmp_path: Path) -> None:
-    """Rewriting a packed shard must keep one row group per episode, not collapse
-    every episode into a single giant row group."""
-    episode_lengths = [4, 6, 5]  # unequal lengths, all in one shard
-    root = _build_packed_dataset(tmp_path / "ds", episode_lengths)
-    shard = root / "data" / "chunk-000" / "file-000.parquet"
-    assert pq.ParquetFile(shard).metadata.num_row_groups == 1, "fixture should start collapsed"
-
-    staging_dir = tmp_path / "stage"
-    for ep in range(len(episode_lengths)):
-        _stage_episode(
-            staging_dir,
-            ep,
-            plan=[
-                {
-                    "role": "assistant",
-                    "content": f"subtask for ep {ep}",
-                    "style": "subtask",
-                    "timestamp": 0.0,
-                    "tool_calls": None,
-                }
-            ],
-        )
-
-    records = list(iter_episodes(root))
-    LanguageColumnsWriter().write_all(records, staging_dir, root)
-
-    # One row group per episode, with row counts matching the episode lengths.
-    md = pq.ParquetFile(shard).metadata
-    assert md.num_row_groups == len(episode_lengths)
-    assert [md.row_group(i).num_rows for i in range(md.num_row_groups)] == episode_lengths
-    # Language columns are still present after the per-episode rewrite.
-    table = pq.read_table(shard)
-    assert "language_persistent" in table.column_names
-    assert "language_events" in table.column_names
-
-
 def test_speech_atom_shape_matches_plan_spec() -> None:
     atom = speech_atom(2.5, "I'm cleaning up!")
     assert atom["role"] == "assistant"

tests/datasets/test_aggregate.py

@@ -32,26 +32,6 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from tests.fixtures.constants import DUMMY_REPO_ID
 
 
-def assert_data_shards_one_row_group_per_episode(root):
-    """Every aggregated DATA shard must have exactly one parquet row group per episode."""
-    import pyarrow.parquet as pq
-
-    shards = sorted((root / "data").rglob("*.parquet"))
-    assert shards, f"no data shards found under {root}/data"
-    n_episodes = 0
-    for shard in shards:
-        pf = pq.ParquetFile(shard)
-        episodes = pf.read(columns=["episode_index"]).column("episode_index").to_pylist()
-        assert pf.metadata.num_row_groups == len(set(episodes)), shard
-        for i in range(pf.metadata.num_row_groups):
-            rg_episodes = set(
-                pf.read_row_group(i, columns=["episode_index"]).column("episode_index").to_pylist()
-            )
-            assert len(rg_episodes) == 1, f"{shard} row group {i} spans episodes {rg_episodes}"
-        n_episodes += len(set(episodes))
-    return n_episodes
-
-
 def assert_episode_and_frame_counts(aggr_ds, expected_episodes, expected_frames):
     """Test that total number of episodes and frames are correctly aggregated."""
     assert aggr_ds.num_episodes == expected_episodes, (
@@ -586,41 +566,6 @@ def assert_image_frames_integrity(aggr_ds, ds_0, ds_1):
             )
 
 
-@pytest.mark.parametrize("use_videos", [True, False], ids=["video", "image"])
-def test_aggregate_one_row_group_per_episode(tmp_path, lerobot_dataset_factory, use_videos):
-    """Aggregated DATA shards keep one row group per episode (not one collapsed group).
-
-    Covers both the non-image (``df.to_parquet``) and image
-    (``to_parquet_with_hf_images``) write branches, including the merge-into-
-    existing-file branch via a low file-size threshold that forces packing.
-    """
-    ds_0 = lerobot_dataset_factory(
-        root=tmp_path / "rg_0",
-        repo_id=f"{DUMMY_REPO_ID}_rg_0",
-        total_episodes=3,
-        total_frames=60,
-        use_videos=use_videos,
-    )
-    ds_1 = lerobot_dataset_factory(
-        root=tmp_path / "rg_1",
-        repo_id=f"{DUMMY_REPO_ID}_rg_1",
-        total_episodes=4,
-        total_frames=80,
-        use_videos=use_videos,
-    )
-
-    aggr_root = tmp_path / "rg_aggr"
-    aggregate_datasets(
-        repo_ids=[ds_0.repo_id, ds_1.repo_id],
-        roots=[ds_0.root, ds_1.root],
-        aggr_repo_id=f"{DUMMY_REPO_ID}_rg_aggr",
-        aggr_root=aggr_root,
-    )
-
-    n_episodes = assert_data_shards_one_row_group_per_episode(aggr_root)
-    assert n_episodes == ds_0.num_episodes + ds_1.num_episodes
-
-
 def test_aggregate_image_datasets(tmp_path, lerobot_dataset_factory):
     """Test aggregation of image-based datasets preserves HuggingFace Image schema.
 

tests/teleoperators/test_rebot_102_leader.py

@@ -18,7 +18,7 @@ from unittest.mock import MagicMock, patch
 
 import pytest
 
-from lerobot.teleoperators.bi_rebot_102_leader import BiRebot102Leader, BiRebot102LeaderConfig
+from lerobot.teleoperators.bi_rebot_102_leader import BiRebotArm102Leader, BiRebotArm102LeaderConfig
 from lerobot.teleoperators.rebot_102_leader import (
     RebotArm102Leader,
     RebotArm102LeaderConfig,
@@ -91,11 +91,11 @@ def test_send_feedback_not_implemented(leader):
 
 def test_bimanual_prefixes_features():
     with patch(f"{_MODULE}.require_package", lambda *a, **kw: None):
-        cfg = BiRebot102LeaderConfig(
+        cfg = BiRebotArm102LeaderConfig(
             left_arm_config=RebotArm102LeaderConfig(port="/dev/null0"),
             right_arm_config=RebotArm102LeaderConfig(port="/dev/null1"),
         )
-        teleop = BiRebot102Leader(cfg)
+        teleop = BiRebotArm102Leader(cfg)
     assert any(k.startswith("left_") for k in teleop.action_features)
     assert any(k.startswith("right_") for k in teleop.action_features)
     assert "left_gripper.pos" in teleop.action_features

tests/training/test_multi_gpu.py

@@ -134,7 +134,7 @@ class TestMultiGPUTraining:
                 f"--output_dir={output_dir}",
                 "--batch_size=4",
                 "--steps=10",
-                "--eval_freq=-1",
+                "--env_eval_freq=-1",
                 "--log_freq=5",
                 "--save_freq=10",
                 "--seed=42",
@@ -177,7 +177,7 @@ class TestMultiGPUTraining:
                 f"--output_dir={output_dir}",
                 "--batch_size=4",
                 "--steps=20",
-                "--eval_freq=-1",
+                "--env_eval_freq=-1",
                 "--log_freq=5",
                 "--save_freq=10",
                 "--seed=42",