# UMI Data with pi0 Relative EE Actions

This guide explains how to prepare a UMI-collected dataset for training a pi0 policy with relative end-effector (EE) actions, and how to run inference with the trained model.

**What we will do:**

1. How to add `observation.state` to an existing UMI LeRobot dataset.
2. How to train pi0 with `use_relative_actions=True`.
3. How to evaluate the trained policy on a real robot.

## Background

[UMI (Universal Manipulation Interface)](https://umi-gripper.github.io) collects manipulation data with hand-held grippers, recovering 6-DoF EE poses via SLAM. UMI datasets stored in LeRobot format already contain `action` (absolute EE pose) and wrist-camera images. To train pi0 with relative actions, we need two additions:

1. **`observation.state`** — the current EE pose the policy conditions on.
2. **Relative action statistics** — so the normalizer sees `(action − state)` distributions.

### Why relative actions?

With relative actions, each action in a chunk is an **offset from the current state** rather than an absolute target:

```
relative_action[i] = absolute_action[t + i] − state[t]
```

This is the representation advocated by UMI (Chi et al., 2024). Compared to absolute actions it removes the need for a consistent global coordinate frame, and compared to delta actions (each step relative to the previous) it avoids error accumulation across the chunk. See the [Action Representations](action_representations) guide for a full comparison.

## State-Action Offset

UMI SLAM produces a single trajectory of EE poses stored as `action`. We derive `observation.state` from the same trajectory with a configurable offset:

```
state[t] = action[t - offset]
```

| Offset | `state[t]`    | Meaning                                                          |
| ------ | ------------- | ---------------------------------------------------------------- |
| 0      | `action[t]`   | State and action are the same pose at time t                     |
| 1      | `action[t-1]` | State is the previous action — where the gripper already arrived |

An offset of 1 is the typical UMI convention: at decision time the "current state" is where the gripper _already is_ (the result of the previous command), and the action is where it should go next. At episode boundaries where `t < offset`, we clamp to `action[0]`.

## Step 1: Add `observation.state`

pi0 with `use_relative_actions=True` needs `observation.state` in the dataset to compute `action - state` on the fly. The script in `examples/umi_pi0_relative_ee/convert_umi_dataset.py` adds it. Edit the constants at the top:

```python
HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"

# Option A: Copy an existing feature as observation.state
STATE_SOURCE_FEATURE = "observation.joints"  # or "observation.pose", etc.

# Option B: Derive from action with offset (set STATE_SOURCE_FEATURE = None)
STATE_SOURCE_FEATURE = None
STATE_ACTION_OFFSET = 1
```

**Choosing the state source:**

- If your dataset already has a feature in the same space as `action` (e.g. `observation.joints` for joint-space actions, or `observation.pose` for EE-space actions), set `STATE_SOURCE_FEATURE` to copy it.
- If your dataset only has a single trajectory (like raw UMI EE data where action = the EE poses), set `STATE_SOURCE_FEATURE = None` and use `STATE_ACTION_OFFSET` to derive state from the action column with a time offset.

`observation.state` **must have the same shape as `action`** — the relative conversion computes `action - state` element-wise.

Then run:

```bash
python examples/umi_pi0_relative_ee/convert_umi_dataset.py
```

<Tip>

If your dataset already has `observation.state`, the script exits early — nothing to do.

</Tip>

## Step 2: Recompute Relative Action Stats

Use the built-in CLI to recompute dataset statistics in relative space:

```bash
lerobot-edit-dataset \
    --repo_id <your_dataset> \
    --operation.type recompute_stats \
    --operation.relative_action true \
    --operation.chunk_size 50 \
    --operation.relative_exclude_joints "['gripper']" \
    --push_to_hub true
```

The `relative_exclude_joints` parameter specifies joints that stay absolute. Gripper commands are typically binary or continuous open/close and don't benefit from relative encoding. Leave it as `"[]"` to convert all dimensions to relative.

## Step 3: Train

No custom training script is needed — standard `lerobot-train` handles everything:

```bash
lerobot-train \
    --dataset.repo_id=<hf_username>/<dataset_repo_id> \
    --policy.type=pi0 \
    --policy.pretrained_path=lerobot/pi0_base \
    --policy.use_relative_actions=true \
    --policy.relative_exclude_joints='["gripper"]'
```

Under the hood, the training pipeline:

- Loads relative action stats from the dataset's `meta/stats.json`.
- Configures `RelativeActionsProcessorStep` in the preprocessor (absolute → relative before normalization).
- The model trains on normalized relative action values.

See the [pi0 documentation](pi0) for all available training options.

## Step 4: Evaluate

The evaluation script in `examples/umi_pi0_relative_ee/evaluate.py` runs inference on a real robot (SO-100 with EE space):

```bash
python examples/umi_pi0_relative_ee/evaluate.py
```

Edit `HF_MODEL_ID`, `HF_DATASET_ID`, and robot configuration at the top of the file.

The inference flow uses pi0's built-in processor pipeline — no custom wrappers needed:

1. **Robot → FK** — Joint positions are converted to EE pose via `ForwardKinematicsJointsToEE`, producing `observation.state`.
2. **Preprocessor** — `RelativeActionsProcessorStep` caches the raw `observation.state`, then `NormalizerProcessorStep` normalizes everything.
3. **pi0 inference** — The model predicts a normalized relative action chunk.
4. **Postprocessor** — `UnnormalizerProcessorStep` unnormalizes, then `AbsoluteActionsProcessorStep` adds the cached state back to get absolute EE targets.
5. **IK → Robot** — `InverseKinematicsEEToJoints` converts absolute EE targets to joint commands.

## How the Pieces Fit Together

```
Training:
  dataset (absolute EE) → RelativeActionsProcessorStep → NormalizerProcessorStep → pi0 model

Inference:
  robot joints → FK → observation.state (absolute EE)
                           ↓
                   RelativeActionsProcessorStep (caches state)
                           ↓
                   NormalizerProcessorStep → pi0 model → relative action chunk
                           ↓
                   UnnormalizerProcessorStep
                           ↓
                   AbsoluteActionsProcessorStep (+ cached state → absolute EE)
                           ↓
                   IK → joint targets → robot
```

## References

- [UMI: Universal Manipulation Interface](https://umi-gripper.github.io) — Chi et al., 2024. Defines relative trajectory actions.
- [Action Representations](action_representations) — LeRobot guide comparing absolute, relative, and delta actions.
- [pi0 documentation](pi0) — Full pi0 configuration including `use_relative_actions`.
- [`examples/so100_to_so100_EE/`](https://github.com/huggingface/lerobot/tree/main/examples/so100_to_so100_EE) — EE-space evaluation example this builds on.