Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('webdataset', {}), NamedSplit('validation'): ('json', {}), NamedSplit('test'): ('webdataset', {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Corrected diverse pushing demonstrations, v2

520 actuator-executed simulation demonstrations: 400 training, 40 validation, 80 test. Four tasks: push the tennis ball left/right of a cube/bowl. Each task has 100 train, 10 validation, 10 in-distribution test and 10 spatial-OOD test episodes. This is a new release; the previous release remains unchanged.

Split Episodes Action steps
train 400 136,439
validation 40 13,738
test 80 27,778

Corrections and diversity

Both directions use a 0.5 s velocity ramp and approximately 40 mm/s cruise. Push endpoints account for reference size, 20 mm required clearance and a 45 mm margin. Measured settled edge clearance is 64.28–65.12 mm. Pushing after success lasts 1.15–1.20 s. The hand lifts away to release contact, and actual visible fingers stay at least 11.04 mm above the table. The 65 mm, 60 g ball uses hollow spherical-shell inertia (2mr²/3); mass/friction/contact parameters remain simulation estimates. Cube width is 39 mm; bowl diameter is 170 mm.

Train layouts cross five reference-location strata with five safe robot-start regions and four repetitions. Initial signed ball/reference offsets range from −60 to +45 mm; spatial-OOD tests use disjoint −95 to −75 mm and +52 to +62 mm intervals. Independent split seeds are recorded. Per task, training appearance quotas are 4 unchanged, 16 camera-only, 16 lighting-only and 64 both. Original reconstruction textures and lighting style are retained, with episode camera and lighting perturbations. design.json, geometry.json and manifests document the sampling and geometry contracts. Corn tasks are excluded.

Data format and timing

Table and wrist RGB images are 224×224 at 20 Hz. NPZ state has eight actual joint positions in radians (six arm plus two gripper joints). actions has six absolute arm joint targets in radians plus a gripper command in 0–255. These are observations of executed MuJoCo physics, not ideal IK samples or real robot recordings. Servo targets are 100 Hz; physics/state safety traces are 500 Hz. There are N actions and N+1 observations; observation i precedes the interval ending at action i. Assemble 20-step/one-second chunks within an episode and mask terminal padding. The root manifest indexes all tasks and splits; use each row's instruction. Fit normalization to train only: supplied norm_stats.json uses the 400 training episodes. Keep validation and test episodes out of training. evaluation_domain distinguishes ID and spatial-OOD test episodes.

Download

from huggingface_hub import snapshot_download
snapshot_download('bx6d/r2s2r-pushing-diverse400-v2', repo_type='dataset', local_dir='pushing-v2')
cd pushing-v2
sha256sum -c SHA256SUMS
for shard in data/*.tar; do tar -xf "$shard"; done

The 28 uncompressed tar shards total 4.73 GiB and extract into tasks/<task>/<split>/episode_XXXXXX/. Episodes contain episode.json, trajectory.npz, render.json, demo.mp4, images/{table,wrist}/*.jpg, and independent physics/RGB proofs. Native NPZ/JPEG/tar format; no LeRobot/Parquet conversion or tabular dataset viewer is configured. shards.json records exact split membership, byte sizes and SHA-256 digests.

Validation and provenance

All 520 episodes passed independent 500 Hz actual-state safety, task success, timing, visibility, diversity and RGB checks. The final audit decoded every one of the 356,950 JPEG images; proof reports are under reports/. Raw episode/proof metadata retain generator-local scene/asset paths as provenance; these paths are not required to load training observations. Simulator assets, model weights, credentials and real-camera reference images are not included. This release does not establish real-robot dynamics or camera calibration. No dataset license has been assigned.

Downloads last month
243

Models trained or fine-tuned on bx6d/r2s2r-pushing-diverse400-v2