SafeControl/UniVTAC-isaac51-shelf-filtered-100

UniVTAC Isaac 5.1 shelf: 100 filtered native RGB-D demonstrations

This public dataset was collected for SafeControl. The current target is 100 accepted RGB-D episodes, and completion requires COMPLETED.json plus accepted_manifest.json with 100 native-rgbd-v1 entries.

Current versus legacy data

• rgbd/hdf5/<source_seed>.hdf5: current RGB-D collection.
• rgbd/attempts/, rgbd/shards/: current attempt records and resumable ledgers.
• hdf5/, attempts/, shards/: preserved legacy RGB-only collection, stopped when camera depth became a requirement. These episodes do not count toward the 100 RGB-D target and must not be mixed into it.
• archive/rgb-only-v1/: the prior experiment, worker, README and stop record.
• validation/rgbd-smoke/: single-episode validation and comparison evidence.
• accepted_manifest.json, metadata.json, SHA256SUMS, COMPLETED.json: final outputs for the current RGB-D collection only.

What is unchanged and what is added

The public upstream isaac51 source is pinned to d541e5568227ca3b66104d294f63c80acad7c52c. All 39 downloaded upstream source/config/instruction files are hash checked and are not edited. Assets and third-party dependencies come from the frozen Isaac Sim 5.1 container. This does not establish bit-exact reproduction of the original HF release.

The generator, physics, PD gains, forced joint-state writes, motion planning, atom transitions, native success checks, tactile settings, 120 Hz physics and 60 Hz saved observations remain unchanged. The sole runtime config override is:

observation_settings.camera=[rgb,depth]

The original camera definitions already initialize RGB and depth outputs. A recording wrapper adds per-frame camera intrinsics, a depth validity mask. It delegates to the upstream observation serializer. No joint/actor trajectory smoothing or relabeling is performed.

The 21 original HDF5 dataset paths remain present. For each of head and wrist, the following fields are added under observation/<camera>/:

• depth: raw float32 image-plane depth in metres, native 480x270 grid.
• depth_valid_mask: bool, finite depth with 0 < z < 100 m.
• intrinsics: per-frame 3x3 camera K on the native grid.

Camera world extrinsics are not provided: the installed camera world-pose query remained stale for the moving wrist. The earlier camera-coordinate pointmap interface requires depth and K, and does not require these extrinsics.

RGB, depth, intrinsics, joints, actors and tactile frames share the same recorded physical-step indices. Depth is obtained from Isaac Sim rendering, not inferred by DA3. Raw invalid depth values are retained with an explicit mask. Depth is not normalized, clipped to 2 m, converted to an image preview, or resampled in the stored HDF5.

The earlier training adapter expects depth in millimetres: convert stored metres by multiplying by 1000 at that interface. RGB/depth must receive the same spatial transform and K must be transformed consistently. Camera head maps to the prior cam_high name; wrist maps to cam_right_wrist. This dataset does not change the policy training code.

Frozen trajectory quality gate

The gate applies to the seven arm joints in radians at 60 Hz; fingers are metres and are excluded from arm derivative limits. See filter_spec.json and shelf_quality_filter.py for exact definitions.

• Native generation success is required.
• Numeric joint/actor values must be finite, physical-step increments equal 2, and all recorded arrays have matching frame counts.
• Position limits allow 0.005 rad numerical tolerance.
• Speed <= 1.02 times the corresponding URDF speed limit.
• Peak finite-difference acceleration <= 25 rad/s^2.
• One-frame backtrack <= 0.003 rad.
• At most two reversals of >= 0.001 rad on the same joint in a 0.2 s window.
• Jerk and +/-0.1 s transition-window metrics are diagnostic.

These unchanged motion gates retained 57 of the 100 audited original episodes. Additional RGB-D integrity checks require both cameras, aligned grids and frame counts, valid calibration, usable depth in every frame and correct validity masks. They do not relax the motion gate. No policy improvement or hardware safety certification has been established.

The new HDF5 schema has 27 dataset paths: the 21 original paths plus depth, validity mask and intrinsics for each of the two cameras.

Seeds and scene coverage

Seeds start at 1000, using eight disjoint streams: (seed - 1000) modulo 8. Seed1000 from the validated RGB-D smoke can count if it passes both gates. Seeds may match a legacy RGB-only attempt, but RGB-D files are independently generated and evaluated; their trajectory need not be bit-identical.

The target coarse scene counts remain 30/34/19/17 for shelf x >=/< 0.9 crossed with bottle y >=/< 0, in the order recorded in experiment.json. Filtering, these quotas, starting seeds, parallel scheduling and a target of 100 accepted episodes are custom protocol choices, not official dataset release settings.

Source: https://github.com/univtac/UniVTAC/tree/d541e5568227ca3b66104d294f63c80acad7c52c Calibration data: https://huggingface.co/datasets/byml/UniVTAC/tree/e1aee7b0c95543b535e0146b2de3ee1bc6ddaabd/isaac51/put_bottle_in_shelf

Collection finished with 100 accepted RGB-D episodes. Remaining episodes were first checkpointed to persistent storage on nebius-uk-south2, then uploaded together. Original source generator, fixed motion filter, camera-depth recording and coverage quotas were preserved. See migrations/persistent-buffer-bulk-upload.json for storage provenance. COMPLETED.json is written only after all 100 remote sizes and SHA-256 values have been verified.