ForceLDM Dataset
Contact-rich manipulation demonstrations with calibrated force, torque, and optical flow.
This dataset accompanies ForceLDM: A Force-aware Latent Dynamics Model for Contact-Rich Manipulation.
- Paper: ForceLDM: A Force-aware Latent Dynamics Model for Contact-Rich Manipulation
- Project page: ForceLDM
- Dataset: PeiYuan001/forceldm-dataset
This dataset contains 474 episodes and 362,339 frames at 30 Hz, covering USB insertion, gear assembly, box flipping, board wiping, VGA insertion, and cucumber peeling. It uses the LeRobot v2.1 directory layout, with synchronized observations and actions in Parquet files and four video streams per episode.
Contents
| Signal | Description |
|---|---|
observation.images.head_camera |
Head-camera RGB video, 640 × 360 |
observation.images.wrist_left_camera |
Wrist-camera RGB video, 640 × 360 |
observation.future_flow.base_0_rgb |
Future optical-flow RGB visualization, 224 × 224 |
observation.future_flow.left_wrist_0_rgb |
Wrist-view future optical-flow RGB visualization, 224 × 224 |
observation.effort |
Calibrated world-frame [Fx, Fy, Fz, Mx, My, Mz], in N and Nm |
observation.state |
27-dimensional robot state; entries [7:13] contain the same calibrated wrench |
action |
Seven-dimensional end-effector pose and gripper action |
| Stage and contact annotations | Original stage, skill-primitive, and contact labels |
All episodes belong to the original training split. No held-out evaluation split is introduced.
Force calibration
The original sensor wrench is corrected using the saved sensor/tool calibration: subtract sensor bias and pose-dependent gravitational force and moment, then rotate into world coordinates using the acquisition driver's convention. TCP orientations are reconstructed from xyz Euler angles in radians. Both force fields are updated consistently.
No additional per-episode taring, smoothing, clipping, or sign flipping is applied. The same saved calibration parameters, including their fitted signs, are used for all episodes; this assumes that the saved sensor/tool calibration remains applicable. Unloaded readings are not forced to zero. Raw demonstrations are not used to re-fit the calibration.
Per-episode and global normalization statistics are recomputed for both changed fields. Actions, poses, timestamps, videos, and original stage/contact labels are preserved. Contact labels have not been re-derived from the calibrated wrench. Optical flow is stored as an RGB visualization, not as calibrated displacement vectors.
Example: episode 83
A recorded USB insertion demonstration with initial contact, withdrawal, and reinsertion. The synchronized replay combines both RGB cameras with calibrated force and torque curves. This is a recorded demonstration, not an autonomous recovery rollout or a probe prediction.
Synchronized video · Six-axis wrench plot · Per-frame wrench and pose CSV
Layout and download
data/chunk-000/episode_XXXXXX.parquet
videos/chunk-000/<video_key>/episode_XXXXXX.mp4
meta/info.json
meta/stats.json
meta/episodes.jsonl
meta/episodes_stats.jsonl
meta/tasks.jsonl
calibration/
examples/episode_000083/
hf download PeiYuan001/forceldm-dataset --repo-type dataset --local-dir ./forceldm-dataset
Use this dataset's updated statistics when training on the calibrated force signals. Models trained with raw sensor wrench require their original preprocessing.
Verification and reproducibility
The conversion script, original compensation source, calibration parameters, per-episode summaries, and validation reports are included under calibration/. All 1,896 video copies were checked against the source using SHA-256. All non-force Parquet fields were checked for equality, and both modified fields and their statistics were verified. The synchronized example video was fully decoded and checked for frame count and rate.
- Downloads last month
- 202