Collection target: 100,000 sequences. This repository is populated incrementally; the target is not a claim that all sequences are already uploaded. Each archive contains 200 sequences. Each sequence contains 52 frames indexed 0–51, at 30 Hz.

The original GS collection controller is retained: movement starts at frame 30 and uses a 15-frame smoothstep interpolation. This is an IK-controlled collection dataset, not checkpoint-driven evaluation. The circular rim stays visible; its original filled inner visual is hidden so the upper camera can see through the opening. The original collision geometry is retained. No hanging net is added. Camera locations remain those of the supplied original collection command.

Sampling

Seed labels 0–9999 repeat across ten rounds. Each sequence uses SeedSequence([global_episode_id % 10000, global_episode_id // 10000, 20260908]), producing reproducible, distinct streams.
The forward catch disk (radius 1.4 m, forward coordinate at least 0.3 m) is mapped into 100 × 100 equal-area cells. This is not uniform sampling in polar radius.
The original IK checks cell centers and four interior points. Feasible cells are allocated balanced sequence counts, then sampled randomly inside. Every actual throw is checked again by the original collection controller. Excluded cells and the precise global episode/cell mapping are in metadata/grid_plan.*.
Launch distance: 3–5 m forward; launch world Z: 0.81–1.0 m; target world Z: 0.9 m; flight duration: 1.5–1.6 s quantized to the original 0.0333333 s physics step. Base landing share: 0.7.
Temporal and launch parameters remain random within their ranges. The equal-area stratification applies to catch location.

Priority tiers

core/shard_XXXXXX.tar is uploaded first. Each contains per-camera HEVC/H.265 MP4, parameters.npz, and labels for 200 sequences. parameters.npz contains all 52 states, actions, frame indices, timestamps and camera transforms. State order is base X/Y/yaw plus six arm joints. action[t] = state[min(t+1, 51)], matching the original action-lookahead setting. Camera intrinsics and collection configuration accompany each shard.

aux/shard_XXXXXX.tar contains losslessly compressed float32 depth, segmentation class IDs, and complete per-frame measured simulation parameters: commanded actions, joint positions/velocities, root/base/ring poses, root velocities, ball position/velocity, ring center/normal and camera transforms, plus throw and IK labels. Sensor-invalid depth values are retained.

Auxiliary uploads begin only after 10,000 core sequences have been uploaded and verified, and core data keeps queue priority thereafter. An absent auxiliary shard may simply still be queued.

Compression and integrity

RGB is encoded as HEVC MP4 (yuv420p, CRF 24, medium preset). If full-image or ball-region checks require it, encoding retries use yuv444p at CRF 20, then CRF 12, then lossless HEVC as a final fallback. RGB-to-YUV conversion still applies, so lossless HEVC is bit-exact in YUV space rather than a promise of identical original RGB bytes. Actual pixel format, CRF, decoded frame counts and sampled PSNR are stored per episode, with encoding_audit.json recording the attempted settings. The recorder retains RGB in memory until the MP4 is decoded and checked, then releases it; no raw RGB files are written to disk. Final states/actions remain numeric arrays, not video-encoded values.

Depth float32 bit patterns are compressed losslessly using temporal XOR, byte shuffling and NPZ DEFLATE. Use depth_codec.decode(depth_xor_shuffled, depth_shape) to recover the exact original array, including invalid-value bit patterns.

Every archive includes a per-file SHA256SUMS. Local archives are removed only after Hugging Face confirms their exact size and SHA-256. Persistent queues retain pending transfers and retry network errors. Collection pauses before consuming the configured disk reserve.

Before upload, archives are checked against every per-file checksum, the complete 200-episode ID range, required camera and parameter files, 52-frame arrays, finite numeric parameters, video quality receipts, and next-frame state/action alignment. The first published core shard was also downloaded from Hugging Face and passed this content audit, alongside its retained auxiliary shard.

Model input

Decode frames [24,27,30] for the current three-observation VLA query at frame 30. Because this follows original collection timing, frame 30 already contains the first small movement. All 52 frames are available for other query times. Use the saved task string, state vectors and camera-key order left, right, upper; apply your checkpoint's own image preprocessing and normalization.

load_episode.py provides a NumPy/PyAV reader. MP4-decoded RGB, rather than deleted pre-encoding pixels, defines the released dataset's observation images. The saved archive layout is explicit NPZ/JSON/MP4, not a drop-in LeRobot folder.

Camera extrinsics are column-vector USD camera-to-world transforms. Convert OpenCV optical coordinates to USD camera coordinates with diag(1,-1,-1,1). Geometry uses metres; joint angles use radians; pose quaternion order is XYZW. The theoretical target is an airborne catch location, not ground impact.

Downloads last month: -

Total file size:

7.01 GB