README.md

---
license: mit
task_categories:
  - robotics
tags:
  - LeRobot
  - SO-101
  - SO-100
  - robot-manipulation
  - teleoperation
  - imitation-learning
  - world-models
  - action-conditioned-video-prediction
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files: data/*/*.parquet
---

# SO-101 Dual-View Teleoperation Dataset for World Models

This dataset contains leader-follower teleoperation recorded with an SO-101
follower arm, six joint-position command/state channels, and two synchronized
RGB cameras. It was collected in two phases: repeated goal-directed
pick-and-place demonstrations for visuomotor policy research, followed by
longer exploratory interaction intended to broaden the state-transition
coverage available to visual world models.

This is the privacy-cleaned release. Identifiable faces in the wrist-camera
stream were blurred without deleting frames or changing video timing. The
recorded Parquet rows, timestamps, commands, and measured follower states are
not shifted or rewritten to compensate for physical control delay.

## What raw means in this release

This is a privacy-cleaned raw sensor release, not a derived policy dataset.

Unchanged recorded data:

- Every Parquet action, measured state, identifier, frame index, and timestamp
- Every fixed-camera frame
- Video frame order, timing, duration, geometry, and codec
- Episode boundaries and task strings

Intentional release additions or modifications:

- Identifiable faces are blurred in five FPV files. Re-encoding changes pixels
  and compressed bytes but preserves every per-frame timestamp.
- Human interventions are described by an exclusion manifest. No source rows
  or frames are deleted or concatenated.
- Repaired descriptive statistics, train-only normalization, deterministic
  episode splits, audit reports, and training adapters are included.

The complete V-JEPA 2-AC procedure is in [TRAINING.md](TRAINING.md). Run its
preflight before submitting a GPU job.

## Dataset summary

| Property | Value |
|---|---:|
| Episodes | 177 |
| Frames | 355,884 |
| Recorded duration | 3.295 hours |
| Recording rate | 30 Hz |
| Robot state dimension | 6 |
| Robot action dimension | 6 |
| Cameras | 2 synchronized RGB streams |
| Video format | AV1, 640 x 480, 30 fps, `yuv420p` |
| Audio | None |
| Depth/force/torque | Not recorded |
| LeRobot format | v3.0 |

All episodes remain addressable through the standard LeRobot `train` split for
compatibility. A deterministic, collection-stratified 141/18/18 episode split
is provided in `meta/cleaning/episode_splits.json`; use that manifest for model
selection and final evaluation rather than splitting frames randomly.

## Collection phases

The two `task_index` values describe collection intent. They are not
success/failure or reward labels.

| `task_index` | Episodes | Frames | Share of frames | Collection intent |
|---:|---:|---:|---:|---|
| 0 | 0-149 | 113,691 | 31.95% | Repeated goal-directed pick-and-place demonstrations originally collected for VLA/policy training |
| 1 | 150-176 | 242,193 | 68.05% | Longer exploratory interaction collected to broaden world-model dynamics coverage |

The task strings stored in `meta/tasks.parquet` are:

```text
0: Grab the red hexagon on the right and place it on the red hexagon on the left.
1: exploratory training for world models
```

Only 27 of the 177 episodes are exploratory, but those episodes are much
longer and therefore contain approximately 68% of all frames. Uniform random
frame sampling will consequently devote approximately 68% of optimization to
the exploratory phase. This is not data corruption; it is a training-sampling
decision.

For action-conditioned world-model experiments, compare at least:

- Natural frame mixture: approximately 32% demonstrations / 68% exploration
- Balanced mixture: 50% demonstrations / 50% exploration
- Task-weighted mixture: 70% demonstrations / 30% exploration

The 70/30 setting is a recommended experiment, not a claim that it is optimal.
Select the mixture using held-out results reported separately for the two
collection phases. Do not delete the unused group to create a mixture; choose
the group and episode hierarchically in the sampler.

Sampling ratio and dataset split are separate choices: use only training
episodes first, then apply the desired demonstration/exploration mixture
inside that set.

## Recording setup

- **Follower:** SO-101 arm (`robot_type: so_follower`).
- **Input device:** paired SO-100/SO-101-style leader arm.
- **Control:** the leader supplies commanded joint-position targets and the
  follower reports its measured joint positions.
- **Rate:** commands, follower state, and both cameras are recorded at 30 Hz.
- **Fixed camera:** `observation.images.left`, a stable global workspace view.
- **Wrist camera:** `observation.images.fpv`, a moving local view near the
  gripper and contact area.

The two camera streams and the numeric rows are synchronized. Matching record
timestamps do not imply an instantaneous mechanical response: the follower
requires time to process and physically execute a leader command.

## Data fields

| Key | Type | Shape | Meaning |
|---|---|---:|---|
| `action` | float32 | `(6,)` | Commanded leader joint-position target sent toward the follower |
| `observation.state` | float32 | `(6,)` | Measured follower joint position |
| `observation.images.left` | video | `(480, 640, 3)` | Fixed third-person RGB camera |
| `observation.images.fpv` | video | `(480, 640, 3)` | Wrist-mounted RGB camera |
| `timestamp` | float32 | `(1,)` | Seconds from the start of the episode |
| `frame_index` | int64 | `(1,)` | Frame index inside the episode |
| `episode_index` | int64 | `(1,)` | Episode identifier |
| `index` | int64 | `(1,)` | Global row identifier |
| `task_index` | int64 | `(1,)` | Collection-phase identifier |

Joint ordering for both `action` and `observation.state`:

```text
shoulder_pan.pos
shoulder_lift.pos
elbow_flex.pos
wrist_flex.pos
wrist_roll.pos
gripper.pos
```

The integer identifiers and timestamps are alignment metadata. They should not
be normalized or supplied as physical action/state values.

## Command-to-motion delay

Cross-correlation between commanded targets and future measured follower state
shows an end-to-end response delay of approximately:

| Joint | Estimated response delay |
|---|---:|
| Shoulder pan | 4 frames / 133 ms |
| Shoulder lift | 4 frames / 133 ms |
| Elbow flex | 5 frames / 167 ms |
| Wrist flex | 4 frames / 133 ms |
| Wrist roll | 4 frames / 133 ms |
| Gripper | Approximately 4 frames / 133 ms, with larger residual error |

This is not evidence that timestamps were recorded incorrectly. It is the
combined delay of leader sensing, processing/communication, follower control,
motor mechanics, and follower-state measurement. The available data does not
contain separate timestamps for each of those stages, so their individual
contributions cannot be isolated retrospectively.

### Recommended action conditioning

Do not modify the raw dataset by shifting action rows. Construct temporal
conditioning windows inside the loader.

For a visual transition sampled at approximately 4 Hz (250 ms):

1. Read the current camera clip and measured follower state.
2. Include approximately five preceding 30 Hz commands to cover the measured
   133-167 ms response delay.
3. Include the seven or eight ordered commands issued over the visual
   transition.
4. Resample the complete interval to a fixed ordered sequence and encode it
   with a learned temporal action projection.
5. Predict the next visual latent. Next measured state prediction can be added
   as an auxiliary experiment, but it is not part of the supplied baseline.

Because `30 / 4` is not an integer, preserve timestamps and use timestamp-based
resampling/masks rather than assuming every visual interval contains exactly
the same number of control samples.

This representation allows the model to learn the delay and joint-dependent
response from data without discarding or fabricating measurements.

The bundled adapter uses 13 ordered samples for each transition. Five source
control frames cover the preceding 167 ms, and the remaining temporal extent
reaches the following visual frame. With six joints, the action projection
receives `13 x 6 = 78` ordered values. The current normalized six-dimensional
measured state is encoded separately.

## Wrist-roll cable constraint

The wrist cannot rotate continuously through 360 degrees because the motor
cable limits the safe route. When attempting to reach an equivalent
orientation near the `-180/+180` representation boundary, the controller can
command a value on the opposite side of that boundary while the physical wrist
remains continuous and then reverses direction.

A representative recorded transition occurs in episode 174 near 288.60 s:

```text
commanded wrist target: -179.824 -> +179.912
measured wrist state:   -174.374 -> -175.604
```

The approximate 360-degree numerical command jump is therefore not an
instantaneous physical rotation.

For learning and planning:

- Treat wrist roll as a bounded, cable-constrained joint.
- Preserve the raw commanded target and measured motor coordinate.
- Provide command/state history, velocity, and direction.
- Do not use shortest-angle wrapping as the sole representation.
- Sine/cosine orientation may be added only as an auxiliary feature; it must
  not replace the cable-aware motor coordinate.
- Enforce measured hardware-safe clockwise and anticlockwise limits in any
  controller or planner. Safe limits are hardware configuration, not values
  inferred solely from this dataset.

## Camera characteristics

The fixed camera provides stable global scene context. The wrist camera gives
useful local contact information and includes deliberate look-around motion,
but it is sometimes partly occupied by the gripper or a nearby workpiece.

Temporary FPV occlusion does not make a frame invalid. Recommended models
should combine:

- Both synchronized views
- Several preceding visual frames
- Explicit camera/view embeddings
- Global fixed-camera context with local FPV contact detail

The supplied V-JEPA 2-AC baseline follows the official DROID loader pattern
and randomly selects one synchronized camera view for each clip. This trains
one view-robust predictor without changing the pretrained encoder input shape.
Explicit simultaneous dual-view fusion is a useful later experiment, not a
feature silently claimed by the baseline.

Normal motion blur is retained because it occurs during real robot motion and
is likely to occur at deployment. A clip should be excluded for visual quality
only when several consecutive frames are corrupted or both views simultaneously
lose the robot, objects, and usable scene context. No such systemic corruption
was found during the current integrity review.

## Privacy cleaning

Identifiable faces were found only in selected wrist-camera videos. The cleaned
release blurs faces in:

```text
file-003.mp4
file-005.mp4
file-007.mp4
file-008.mp4
file-010.mp4
```

The cleaning process preserves each affected video's AV1 codec, 640 x 480
resolution, 30 fps rate, `yuv420p` pixel format, duration, frame order, and
exact frame count. The other eight wrist-camera files remain byte-identical to
their source versions. All affected files passed complete decode validation.

An additional manual review confirmed and blurred a face in `file-005.mp4`
from approximately `00:05:01.35` through `00:05:02.92`. Fast-rotation detector
gaps at the entrance and top-edge exit are covered by reviewed manual boxes.
Exact curated windows are recorded in
`meta/cleaning/privacy_blur_windows.json`.

Blurring changes decoded pixels and therefore necessarily changes compressed
file bytes. It does not change video/action/state alignment.

## Human resets and external intervention

Some collection intervals contain a human hand resetting workpieces, and a few
contain possible manual contact with the follower. Such motion is exogenous:
the visual scene or measured follower state may change without being explained
by the recorded robot command.

Do not remove those frames from only one video and concatenate the remaining
sections. That would create a false object teleportation. Instead:

- Keep source videos and Parquet rows unchanged.
- Exclude synchronized video, action, and state intervals together.
- Prevent visual clips and action-history windows from crossing an exclusion.
- Split a logical episode around a reset that occurs inside an episode.
- Add a short guard interval for hand entry, object settling, and hand exit.

Most reviewed resets occur at existing episode boundaries, which already stop
ordinary episode-aware samplers from crossing them. Resets inside longer
episodes still require an explicit exclusion manifest.

The release provides `meta/cleaning/intervention_exclusions.csv`. It contains
17 reviewed source-video intervals mapped into 27 episode-local ranges because
some intervals cross episode boundaries. Fifteen confirmed interventions are
active exclusions; two uncertain human-near-robot intervals remain included
and are marked `reviewed_unconfirmed`. Active intervals include a 0.5-second
guard on both sides. Apply the manifest as a clip-validity mask; it does not
physically remove or concatenate frames.

Examples visible in `observation.images.left`:

| Video | Source time | What is visible |
|---|---:|---|
| `file-000.mp4` | `00:04:21.5-00:04:23.0` | A hand repositions a red workpiece in episode 8 |
| `file-001.mp4` | `00:03:19.0-00:03:22.0` | Two hands reset the red blocks across episodes 19 and 20 |
| `file-021.mp4` | `00:03:27.5-00:03:30.0` | A hand relocates a red block inside exploratory episode 163 |

Exclusion changes which candidate clips may be sampled, not the stored data.
For a candidate start time, the loader considers its complete visual context,
preceding command history, and prediction target. If any part overlaps an
active exclusion, that start is invalid. The sampler selects another valid
start without joining the clean regions on either side. The released 70/30
training configuration contains 269,278 valid eight-frame clip starts.

## Normalization

Normalize only continuous physical features used by the model:

- Six commanded joint targets in `action`
- Six measured joint positions in `observation.state`

Do not normalize `episode_index`, `task_index`, `frame_index`, global `index`,
or timestamps. Compute continuous-feature statistics from the cleaned training
episodes only, after exclusions and train/validation/test assignment, to avoid
validation leakage. Apply the same stored training statistics to validation,
test, and deployment inputs.

Normalization is a reversible affine transformation; it does not delete motion
information. It improves conditioning so joints with larger numerical ranges
do not dominate optimization.

Use `meta/cleaning/normalization_stats.json` for the supplied episode split.
It is computed exactly from 277,478 retained training rows after active
intervention exclusions. `meta/stats.json` is separately repaired as
full-dataset descriptive metadata for LeRobot compatibility; it must not be
substituted for train-only normalization in a controlled evaluation.

## Integrity review

The following checks were performed locally:

- All 177 episode metadata entries agree with Parquet row ranges.
- Global row indices are consecutive.
- Episode timestamps are monotonic.
- Actions and states contain no NaN or infinite values.
- Both camera streams contain 355,884 frames.
- All reviewed videos decode successfully.
- A systematic 1 Hz motion/contrast scan found zero samples where both views
  simultaneously met the conservative unusable-frame criteria.
- Numeric `episode_index` values span 0-176 and `task_index` values span 0-1.
- Privacy-modified videos preserve their original frame counts and stream
  geometry.
- Per-frame presentation timestamps for all five privacy-modified FPV files
  match their original counterparts with measured maximum drift of 0 seconds.

The source release's `meta/stats.json` contained incorrect identifier
aggregates (`episode_index`, `task_index`, and global `index`) and invalidly
small camera standard deviations, even though the underlying Parquet/video
data were sound. The cleaned release repairs those values: numeric statistics
use all exact rows and camera statistics use the documented systematic 1 Hz
sample in `meta/cleaning/video_audit.json`. Identifiers remain metadata and
must not be normalized.

## Reproducibility files

| File | Purpose |
|---|---|
| `meta/cleaning/episode_splits.json` | Deterministic episode-level train/validation/test assignment |
| `meta/cleaning/intervention_exclusions.csv` | Confirmed and reviewed human-intervention intervals in episode-local time |
| `meta/cleaning/normalization_stats.json` | Exact train-only action/state normalization after exclusions |
| `meta/cleaning/video_audit.json` | Cleaned-video frame counts, sampled pixel statistics, and quality triage |
| `meta/cleaning/privacy_blur_windows.json` | Curated face-blur windows and reviewed manual gap coverage |
| `meta/cleaning/alignment_validation.json` | Full cross-modal alignment, immutable-asset, split, exclusion, and normalization audit |
| `meta/cleaning/README.md` | Artifact semantics and regeneration commands |
| `TRAINING.md` | Exact environment, preflight, patch, smoke-test, and distributed-training instructions |
| `training/vjepa2-so101.patch` | Adapter for Meta V-JEPA 2 revision `204698b45b3712590f06245fbfba32d3be539812` |
| `training/preflight.py` | Local frame, split, exclusion, action-window, and sampler verification |
| `training/configure_vjepa2.py` | Writes concrete training configs without hand-editing dataset or checkpoint paths |

## Intended uses

- Action-conditioned latent video prediction
- Dual-view visual world models
- Robot representation learning
- Visuomotor policy and VLA experiments
- Comparing temporal action-conditioning designs
- Offline model-predictive-control research after adding a planner and safety
  constraints

A world model trained on this dataset is not by itself a deployable policy. A
robot deployment still needs action selection/planning, joint and cable safety
limits, collision handling, and hardware supervision.

## Limitations

- Single robot setup and workspace
- Limited visual/environment diversity
- No depth, force, or torque observations
- No explicit success/failure or reward annotations; these are unnecessary for
  dynamics learning but limit reward/value learning
- Few externally caused recovery or failure examples
- End-to-end delay is measured, but its communication/mechanical components are
  not separately timestamped
- Exact safe wrist cable limits must come from the robot/controller setup
- FPV frames can contain expected contact occlusion and motion blur
- Sampling mixture and held-out split must be selected for the downstream
  experiment

## Loading with LeRobot

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("Samisaliveagain/so101_wm")

print(dataset.num_episodes)
print(dataset.num_frames)
sample = dataset[0]
print(sample["action"].shape)
print(sample["observation.state"].shape)
```

For action-conditioned world-model training, use an episode-aware custom
sampler that applies the provided exclusions and constructs timestamp-based
command-history windows. A plain single-frame `action[t] -> video[t+1]` pairing
does not represent the measured follower delay adequately.

## V-JEPA 2 action-conditioned training

The directly supported starting point is Meta's V-JEPA 2 ViT-g encoder and
action-conditioned predictor structure. The official AC implementation expects
seven-dimensional Cartesian DROID trajectories, so it cannot consume this
six-dimensional joint dataset without adaptation. The supplied pinned patch:

- Reads LeRobot v3 directly
- Applies the released splits and intervention mask
- Implements hierarchical 70/30 sampling
- Builds ordered 30 Hz command-history windows for each 4 Hz transition
- Separates the 78-dimensional command projection from the six-dimensional
  measured-state projection
- Preserves the wrist-roll motor coordinate without circular wrapping

Meta's repository includes V-JEPA 2.1 visual backbones, but the pinned official
code does not contain a V-JEPA 2.1 action-conditioned post-training recipe.
This release therefore provides an exact V-JEPA 2-AC baseline first. A 2.1 AC
adaptation must be labeled experimental and separately validate its 384-pixel
encoder, checkpoint layout, tokenizer, and predictor compatibility.

Follow [TRAINING.md](TRAINING.md) exactly. The short form is:

```bash
python training/preflight.py --dataset-root .
git clone https://github.com/facebookresearch/vjepa2.git
cd vjepa2
git checkout 204698b45b3712590f06245fbfba32d3be539812
git apply /path/to/so101_wm/training/vjepa2-so101.patch
python tests/test_so101_index.py
```

Do not deploy a trained predictor directly as a robot policy. Planning,
hardware joint limits, wrist cable limits, collision checks, and supervised
robot testing remain separate requirements.

## Contributors

The raw recordings were collected collaboratively by
[Samisaliveagain](https://huggingface.co/Samisaliveagain) and
[Shubham](https://huggingface.co/shubham4413). This release is maintained and
published from the `Samisaliveagain` account.

## Citation

```bibtex
@misc{so101wm2026,
  title  = {SO-101 Dual-View Teleoperation Dataset for World Models},
  author = {Samisaliveagain and Shubham},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Samisaliveagain/so101_wm}
}
```

## License

MIT