lerobot/docs/source/sarm.mdx

# SARM: Stage-Aware Reward Modeling

SARM (Stage-Aware Reward Modeling) is a video-based reward modeling framework for long-horizon robot manipulation tasks. This guide covers how to train SARM reward models and optionally use them with Reward-Aligned Behavior Cloning (RA-BC).

**Paper**: [SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation](https://arxiv.org/abs/2509.25358)

<img
  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-sarm.png"
  alt="An overview of SARM"
  width="80%"
/>

## Why Reward Models?

Standard behavior cloning treats all demonstration frames equally, but real-world robot datasets are messy. They contain hesitations, corrections, and variable-quality trajectories. Reward models solve this by learning a generalizable notion of **task progress** from demonstrations: given video frames and a task description, they predict how close the robot is to completing the task (0→1). This learned "progress signal" can be used in multiple ways, two promising applications are: (1) **weighted imitation learning** (RA-BC), where high-progress frames receive more weight during policy training, and (2) **reinforcement learning**, where the reward model provides dense rewards for online or offline policy improvement.

## Overview

SARM has following features:

1. **Stage-aware architecture**: Jointly predicts the high-level task stage and fine-grained progress within each stage
2. **Subtask annotations**: Uses natural language subtask annotations to derive consistent progress labels
3. **Temporal proportions**: Computes dataset-level priors (α̅\_k) for each subtask to normalize progress across variable-length demonstrations

SARM trains on a compact **stage+tau** target for each frame:

- **stage**: integer stage index `k ∈ {0, ..., K-1}`
- **τ (tau)**: within-stage progress `τ ∈ [0, 1]`
- **target encoding**: `y = k + τ` (this is what the dataset processor produces)

At inference time (and in downstream RA-BC), SARM converts the raw `k + τ` value into a **normalized progress** in `[0, 1]` using dataset-level **temporal proportions** `α̅_k` (stored in `meta/temporal_proportions_*.json`).

This matches **Formula (2)** from the paper:

```
progress_t = P_{k-1} + α̅_k × τ_t
```

Where:

- `τ_t = (t - s_k) / (e_k - s_k)` is within-subtask normalized time
- `P_{k-1}` is cumulative prior (sum of previous subtask proportions)
- `α̅_k` is the temporal proportion for subtask k

This ensures identical task states map to consistent progress values, even across demonstrations of different lengths.

## Inputs and Targets (What the new code expects)

SARM is trained through its processor (`src/lerobot/rewards/sarm/processor_sarm.py`), which:

- **Encodes** images and task text with CLIP (ViT-B/32) into `video_features` and `text_features`
- **Pads/truncates** robot state into `state_features` (up to `max_state_dim`)
- **Builds targets** as `sparse_targets` (and `dense_targets` in `dense_only`/`dual`) using the stage+tau encoding `y = k + τ`
- **Masks rewind frames** using a per-sample `lengths` tensor (rewind is a training-time augmentation)

At minimum, each training sample needs:

- `task` (string): task description
- `policy.image_key` images and `policy.state_key` states from the dataset

---

## Annotation Modes

You can choose from **3 annotation modes** that determine how progress labels are computed:

| Mode           | Annotations Required | Heads                        | Use Case                                                     |
| -------------- | -------------------- | ---------------------------- | ------------------------------------------------------------ |
| `single_stage` | None                 | Sparse only                  | Simple tasks, quick experiments, no VLM needed               |
| `dense_only`   | Dense (VLM)          | Dual (sparse auto-generated) | Detailed subtask tracking without defining high-level stages |
| `dual`         | Sparse + Dense (VLM) | Dual                         | Full SARM paper setup with both granularities                |

### Mode Details

<hfoptions id="mode_explanation">
<hfoption id="single_stage">

**No annotations required.** The entire episode is treated as a single stage called `"task"`, and progress is linear from 0 to 1 over the episode duration.

- **Sparse head**: 1 stage ("task"), linear progress
- **Dense head**: Not used
- **Best for**: Simple tasks, quick experiments, or when VLM annotation is not available

## Set Up Your Environment

1. Install LeRobot by following our [Installation Guide](./installation).
2. Install SARM dependencies by running:

```bash
pip install -e ".[sarm]"
```

Workflow:

```
1. Train SARM → 2. Visualize predictions → 3. (Optional) Train policy with RA-BC
```

</hfoption>
<hfoption id="dense_only">

**Only dense (fine-grained) annotations from a VLM.** The sparse head automatically uses a single `"task"` stage covering the full episode, while the dense head learns detailed subtask progression.

- **Sparse head**: 1 stage ("task"), linear progress (auto-generated)
- **Dense head**: Multiple fine-grained stages from VLM annotations
- **Best for**: When you want detailed subtask tracking but don't need to define high-level stages

Workflow:

```
1. Annotate (dense) → 2. Verify → 3. Train SARM → 4. Visualize → 5. (Optional) Train policy with RA-BC
```

</hfoption>
<hfoption id="dual">

**Both sparse and dense annotations from VLM.** Full dual-head mode as described in the SARM paper, with both high-level (sparse) and fine-grained (dense) stage predictions.

- **Sparse head**: High-level stages from VLM annotations
- **Dense head**: Fine-grained stages from VLM annotations
- **Best for**: Complex multi-stage tasks where both granularities are useful

Workflow:

```
1. Annotate (sparse+dense) → 2. Verify → 3. Train SARM → 4. Visualize → 5. (Optional) Train policy with RA-BC
```

</hfoption>
</hfoptions>

---

## Step 1: Subtask Annotation

<hfoptions id="annotation_mode">
<hfoption id="single_stage">

**No annotation required!** Skip this step entirely. The model will use the episode's task description and compute linear progress automatically.

</hfoption>
<hfoption id="dense_only">

Generate **dense (fine-grained) annotations only** using a VLM. The sparse stage will be auto-generated.

```bash
python src/lerobot/data_processing/sarm_annotations/subtask_annotation.py \
  --repo-id your-username/your-dataset \
  --dense-only \
  --dense-subtasks "Bring robot arms up from starting position,Grab near side and do 1st fold,Grab side and do 2nd fold,Grab side and do 3rd fold to finish folding" \
  --video-key observation.images.base \
  --num-workers 4 \
  --push-to-hub
```

**What gets saved:**

- `meta/temporal_proportions_sparse.json` - Auto-generated sparse proportions (`{"task": 1.0}`)
- `meta/temporal_proportions_dense.json` - Dense temporal proportions
- Per-episode columns in `episodes/*.parquet`:
  - `dense_subtask_names`, `dense_subtask_start_frames`, `dense_subtask_end_frames`
  - (also time-based columns: `dense_subtask_start_times`, `dense_subtask_end_times`)

</hfoption>
<hfoption id="dual">

Generate **both sparse (high-level) and dense (fine-grained) annotations** using a VLM.

```bash
python src/lerobot/data_processing/sarm_annotations/subtask_annotation.py \
  --repo-id your-username/your-dataset \
  --sparse-subtasks "Bring arms up from starting position,Fold the towel (3 folds in total)" \
  --dense-subtasks "Bring robot arms up from starting position,Grab near side and do 1st fold,Grab side and do 2nd fold,Grab side and do 3rd fold to finish folding" \
  --video-key observation.images.base \
  --num-workers 4 \
  --push-to-hub
```

**What gets saved:**

- `meta/temporal_proportions_sparse.json` - Sparse temporal proportions
- `meta/temporal_proportions_dense.json` - Dense temporal proportions
- Per-episode columns in `episodes/*.parquet`:
  - `sparse_subtask_names`, `sparse_subtask_start_frames`, `sparse_subtask_end_frames`
  - `dense_subtask_names`, `dense_subtask_start_frames`, `dense_subtask_end_frames`
  - (also time-based columns: `*_subtask_start_times`, `*_subtask_end_times`)

</hfoption>
</hfoptions>

### Annotation Arguments

| Argument               | Description                                                                     |
| ---------------------- | ------------------------------------------------------------------------------- |
| `--repo-id`            | HuggingFace dataset repository ID                                               |
| `--sparse-subtasks`    | Comma-separated list of high-level subtask names                                |
| `--dense-subtasks`     | Comma-separated list of fine-grained subtask names                              |
| `--dense-only`         | Generate only dense annotations (auto-creates sparse "task" stage)              |
| `--video-key`          | Camera/video key to use (e.g., `observation.images.top`)                        |
| `--num-workers`        | Number of parallel GPU workers (default: 1)                                     |
| `--episodes`           | Specific episode indices to annotate (default: all)                             |
| `--skip-existing`      | Skip episodes that already have annotations                                     |
| `--model`              | VLM model (default: `Qwen/Qwen3-VL-30B-A3B-Instruct`)                           |
| `--num-visualizations` | Number of episodes to visualize after annotation (default: 5, set to 0 to skip) |

> **Note**: After annotation completes, 5 episodes are automatically visualized by default. Use `--num-visualizations 0` to skip this step.

---

## Step 2: Verify Annotations

<hfoptions id="verify_mode">
<hfoption id="single_stage">

**No verification needed!** Skip this step.

</hfoption>
<hfoption id="dense_only">

Visualize annotations using the `--visualize-only` flag:

```bash
python src/lerobot/data_processing/sarm_annotations/subtask_annotation.py \
  --repo-id your-username/your-dataset \
  --visualize-only \
  --visualize-type dense \
  --num-visualizations 5 \
  --video-key observation.images.base \
  --output-dir ./subtask_viz
```

</hfoption>
<hfoption id="dual">

Visualize annotations using the `--visualize-only` flag:

```bash
python src/lerobot/data_processing/sarm_annotations/subtask_annotation.py \
  --repo-id your-username/your-dataset \
  --visualize-only \
  --visualize-type both \
  --num-visualizations 5 \
  --video-key observation.images.base \
  --output-dir ./subtask_viz
```

</hfoption>
</hfoptions>

This generates visualizations showing video frames with subtask boundaries overlaid and timeline of subtasks.

### Visualization Arguments

| Argument               | Description                                                    |
| ---------------------- | -------------------------------------------------------------- |
| `--visualize-only`     | Only visualize existing annotations (no generation)            |
| `--num-visualizations` | Number of episodes to visualize (default: 5)                   |
| `--visualize-type`     | Type of annotations to visualize: `sparse`, `dense`, or `both` |

**Tip**: If annotations are inaccurate, adjust your subtask descriptions to be more specific and re-run.

---

## Step 3: Train SARM

<hfoptions id="train_mode">
<hfoption id="single_stage">

Train with **no annotations** - uses linear progress from 0 to 1:

```bash
lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=sarm \
  --policy.annotation_mode=single_stage \
  --policy.image_key=observation.images.base \
  --output_dir=outputs/train/sarm_single \
  --batch_size=32 \
  --steps=5000 \
  --wandb.enable=true \
  --wandb.project=sarm \
  --policy.repo_id=your-username/your-model-name
```

</hfoption>
<hfoption id="dense_only">

Train with **dense annotations only** (sparse auto-generated):

```bash
lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=sarm \
  --policy.annotation_mode=dense_only \
  --policy.image_key=observation.images.base \
  --output_dir=outputs/train/sarm_dense \
  --batch_size=32 \
  --steps=5000 \
  --wandb.enable=true \
  --wandb.project=sarm \
  --policy.repo_id=your-username/your-model-name
```

</hfoption>
<hfoption id="dual">

Train with **both sparse and dense annotations**:

```bash
lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=sarm \
  --policy.annotation_mode=dual \
  --policy.image_key=observation.images.base \
  --output_dir=outputs/train/sarm_dual \
  --batch_size=32 \
  --steps=5000 \
  --wandb.enable=true \
  --wandb.project=sarm \
  --policy.repo_id=your-username/your-model-name
```

</hfoption>
</hfoptions>

### Multi-GPU Training

Add `accelerate launch --multi_gpu --num_processes=4` to use multiple GPUs for training.

### Training Arguments

| Argument                   | Description                                                       | Default                  |
| -------------------------- | ----------------------------------------------------------------- | ------------------------ |
| `--policy.annotation_mode` | `single_stage`, `dense_only`, or `dual`                           | `single_stage`           |
| `--policy.image_key`       | Camera key for images                                             | `observation.images.top` |
| `--policy.state_key`       | Key for joint states                                              | `observation.state`      |
| `--policy.n_obs_steps`     | Observation history steps (total obs frames = `n_obs_steps + 1`)  | `8`                      |
| `--policy.frame_gap`       | Gap (in frames) between sampled observations (at 30 fps: 30 ≈ 1s) | `30`                     |

---

## Step 4: Visualize Predictions

Use `compute_rabc_weights.py` with `--visualize-only` to visualize model predictions (and, if available, annotation-derived targets) without writing a parquet file.

<hfoptions id="viz_mode">
<hfoption id="single_stage">

```bash
python -m lerobot.rewards.sarm.compute_rabc_weights \
  --dataset-repo-id your-username/your-dataset \
  --reward-model-path your-username/sarm-model \
  --visualize-only \
  --num-visualizations 5 \
  --head-mode sparse \
  --output-dir ./sarm_viz
```

</hfoption>
<hfoption id="dense_only">

```bash
python -m lerobot.rewards.sarm.compute_rabc_weights \
  --dataset-repo-id your-username/your-dataset \
  --reward-model-path your-username/sarm-model \
  --visualize-only \
  --num-visualizations 5 \
  --head-mode dense \
  --output-dir ./sarm_viz
```

</hfoption>
<hfoption id="dual">

```bash
python -m lerobot.rewards.sarm.compute_rabc_weights \
  --dataset-repo-id your-username/your-dataset \
  --reward-model-path your-username/sarm-model \
  --visualize-only \
  --num-visualizations 5 \
  --head-mode both \
  --output-dir ./sarm_viz
```

</hfoption>
</hfoptions>

The visualization shows:

- **Progress plot**: Predicted progress (and optional annotation-derived “GT” when available and `--stride 1`)
- **Stage probabilities**: Stacked area plot of predicted stage probabilities
- **Sample frames**: Key frames from the episode with progress/stage labels

### Visualization Arguments

| Argument               | Description                                               |
| ---------------------- | --------------------------------------------------------- |
| `--visualize-only`     | Only visualize predictions (no RABC computation)          |
| `--num-visualizations` | Number of episodes to visualize (default: 5)              |
| `--head-mode`          | SARM head to use: `sparse`, `dense`, or `both`            |
| `--stride`             | Compute every N frames, interpolate the rest (default: 1) |

---

## Step 5 (Optional): Train Policy with RA-BC

Reward-Aligned Behavior Cloning (RA-BC) uses the trained SARM model to weight training samples based on predicted progress improvement. This requires two steps:

1. **Precompute progress values** for all frames using the trained SARM model
2. **Train policy** with RA-BC weighting using the precomputed values

### How RA-BC Works

For each training sample, RA-BC computes the progress delta:

```
r_i = φ(o_{t+Δ}) - φ(o_t)
```

Where `φ` is the SARM progress prediction and `Δ` is the policy's `chunk_size`. Samples with positive progress (good demonstrations) get higher weights, while samples with negative or zero progress get down-weighted.

The weighting follows **Equations 8-9** from the paper:

- **Soft weight**: `w̃_i = clip((r_i − (μ − 2σ)) / (4σ + ε), 0, 1)`
- **Final weight**: `w_i = 𝟙{r_i > κ} + 𝟙{0 ≤ r_i ≤ κ} × w̃_i`

### Step 5a: Compute SARM Progress Values

First, run the SARM model on all frames in your dataset to compute progress values:

```bash
python -m lerobot.rewards.sarm.compute_rabc_weights \
  --dataset-repo-id your-username/your-dataset \
  --reward-model-path your-username/sarm-model \
  --head-mode sparse \
  --num-visualizations 5 \
  --push-to-hub
```

This script:

- Processes all frames and computes progress values
- Saves progress values to a parquet file next to the dataset on disk (defaults to `<dataset_root>/sarm_progress.parquet`)
- Generates visualizations of the first N episodes (default: 5)

**Arguments:**

| Argument               | Description                                                    | Default    |
| ---------------------- | -------------------------------------------------------------- | ---------- |
| `--reward-model-path`  | Path to trained SARM model                                     | (required) |
| `--head-mode`          | SARM head to use: `sparse`, `dense`, or `both`                 | `sparse`   |
| `--device`             | Device for inference                                           | `cuda`     |
| `--visualize-only`     | Only visualize predictions (no RA-BC computation)              | `false`    |
| `--num-visualizations` | Number of episodes to visualize (default: 5, set to 0 to skip) | `5`        |

**Output format** (`sarm_progress.parquet`):

| Column            | Description                                    |
| ----------------- | ---------------------------------------------- |
| `index`           | Global frame index in dataset                  |
| `episode_index`   | Episode number                                 |
| `frame_index`     | Local frame index within episode               |
| `progress_sparse` | Sparse head progress value [0, 1]              |
| `progress_dense`  | Dense head progress value [0, 1] (if computed) |

### Step 5b: Train Policy with RA-BC

Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`) if not explicitly provided. Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:

```bash
lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=pi0 \
  --sample_weighting.type=rabc \
  --sample_weighting.head_mode=sparse \
  --sample_weighting.kappa=0.01 \
  --output_dir=outputs/train/policy_rabc \
  --batch_size=32 \
  --steps=40000
```

The training script automatically:

- Loads the precomputed progress values from the parquet file
- Uses the policy's `chunk_size` to compute progress deltas (Δ)
- Computes sample weights based on progress improvement
- Applies weighted loss during training

**RA-BC Arguments:**

| Argument                           | Description                                            | Default                 |
| ---------------------------------- | ------------------------------------------------------ | ----------------------- |
| `--sample_weighting.type`          | Weighting strategy type (`rabc` or `uniform`)          | `rabc`                  |
| `--sample_weighting.progress_path` | Path to progress parquet file                          | `sarm_progress.parquet` |
| `--sample_weighting.head_mode`     | Which SARM head's progress to use: `sparse` or `dense` | `sparse`                |
| `--sample_weighting.kappa`         | Threshold κ for high-quality samples                   | `0.01`                  |
| `--sample_weighting.epsilon`       | Small constant for numerical stability                 | `1e-6`                  |

### Tuning RA-BC Kappa

The `kappa` parameter is the threshold that determines which samples get full weight (w=1). Understanding how to tune it is critical for RA-BC to work effectively.

**How the weighting works:**

| Condition           | Weight                  |
| ------------------- | ----------------------- |
| `delta > kappa`     | 1.0 (hard threshold)    |
| `0 ≤ delta ≤ kappa` | Soft weight from Eq. 8  |
| `delta < 0`         | 0.0 (negative progress) |

**Diagnosing kappa issues:**

Monitor these WandB metrics during training:

| Metric                        | Healthy Range | Problem Indicator         |
| ----------------------------- | ------------- | ------------------------- |
| `sample_weight_mean_weight`   | 0.3 - 0.8     | ≈ 1.0 means kappa too low |
| `sample_weighting/delta_mean` | > 0           | Should be positive        |
| `sample_weighting/delta_std`  | > 0           | Variance in data quality  |

**If `sample_weight_mean_weight ≈ 1.0`:** Your kappa is too low. Most samples have `delta > kappa` and bypass the soft-weighting entirely. RA-BC becomes equivalent to vanilla BC.

**Setting kappa based on your data:**

The default `kappa=0.01` was tuned for the paper's T-shirt folding task (~90s episodes at 30fps). For your dataset, check the logged `sample_weighting/delta_mean` and `sample_weighting/delta_std`:

```
# If delta_mean ≈ 0.03 and delta_std ≈ 0.02:
# Most deltas fall in range [0.01, 0.05]

# Option 1: Set kappa = delta_mean (medium selectivity)
--sample_weighting.kappa=0.03

# Option 2: Set kappa = delta_mean + delta_std (high selectivity)
--sample_weighting.kappa=0.05

# Option 3: Set kappa = delta_mean + 2*delta_std (very selective)
--sample_weighting.kappa=0.07
```

**When RA-BC may not help:**

If your dataset is already high quality (consistent progress across all demonstrations), RA-BC won't provide much benefit since there's nothing to filter.

### Multi-GPU Training with RA-BC

```bash
accelerate launch \
  --multi_gpu \
  --num_processes=4 \
  src/lerobot/scripts/lerobot_train.py \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=pi0 \
  --sample_weighting.type=rabc \
  --sample_weighting.kappa=0.01 \
  --output_dir=outputs/train/policy_rabc \
  --batch_size=32 \
  --steps=40000
```

---

## Tips & Best Practices

### Choosing a Mode

- **Start with `single_stage`** for quick experiments - no annotation overhead
- Use **`dense_only`** when you want detailed progress tracking but tasks don't have clear high-level stages
- Use **`dual`** for complex tasks where both coarse and fine-grained progress is meaningful

### Annotation Quality

1. **Be specific with subtask names**: Instead of "fold", use "grab near side and fold toward center"
2. **Verify with visualization**: Always check a few episodes before training
3. **Consistent naming**: Use the same subtask names across all episodes

### RA-BC

1. **Train SARM first**: RA-BC quality depends entirely on SARM quality
2. **Monitor `sample_weight_mean_weight`**: If it's ≈ 1.0, increase kappa (see [Tuning RA-BC Kappa](#tuning-ra-bc-kappa))

---

## Citation

```bibtex
@article{chen2025sarm,
  title={SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation},
  author={Chen, Qianzhong and Yu, Justin and Schwager, Mac and Abbeel, Pieter and Shentu, Yide and Wu, Philipp},
  journal={arXiv preprint arXiv:2509.25358},
  year={2025}
}
```