G1 full-body teleoperation

30 episodes, 30 Hz, 31 physical state/action dimensions: legs 12, waist 3, arms 14, grippers 2. Three independent 224×224 camera streams: head, left_wrist, right_wrist. Videos use standard browser-compatible H.264 High, YUV420P, BT.709, CRF 1. Original recordings are preserved independently.

Corrected measured state and image validity

Measured full-body state comes from recorded data.json:states.body.qpos, aligned using states.body.received_monotonic_timestamp. Snapshot age must be at most 100 ms; interpolation never bridges measurement gaps above 100 ms. The archival raw DDS subscriber can lag/batch messages; its callback receive time is not used as the sole full-body measurement timeline.

There are 27,709 timeline positions and 27,320 valid training positions. sample_valid excludes actual state/action/image gaps only. Camera frames are always retained independently of full-body validity; missing joint supervision never erases an available image. Truly missing, undecodable, or 100 ms or more distant camera frames are black placeholders. Black borders also come from aspect-ratio-preserving resizing. observation.state_valid and observation.images_is_pad describe their respective sources independently. NaN is used only for unavailable supervision. Respect action_valid_length and action_chunk_mask at true gaps and episode boundaries.

Camera selection uses original camera_frame_raw.jsonl receiver events and their exact video indexes, independently for head/left/right. The uniform 30 Hz numeric clock and original 250→60→30 SDK action resampling are preserved. camera_time_offset_s stores each actual receiver timestamp minus the numeric sample timestamp; camera_valid gives validity independently per camera. Receiver timestamps are not hardware exposure timestamps. Compressed meta/camera_source_frames/ files preserve exact source-video/frame references.

Actions, training and splits

Arm targets are actual SDK commands resampled 250→60→30 Hz. Right gripper uses recorded discrete intent mapped to physical/native position. Legs, waist and uncontrolled left gripper use next measured posture as trajectory supervision; these are not published motor commands. meta/schema.json defines ordered axes, units, label provenance and active axes.

meta/splits.json preserves validation source episodes 10,20,30; the other 27 are training. info.json train range describes repository indexing. Normalization in meta/normalization.json is fit on training episodes only. meta/initial_states/ preserves confirmed stance snapshots. Standard episode indices are zero-based, with original 1-based numbers retained as a column.

This revision supersedes an incorrect state-source export that retained only 8,841 valid positions and used browser-incompatible RGB H.264. That earlier export and its derived training run are not the intended training dataset.

Original full-resolution recordings are retained locally and on the collection/training hosts. This public repository retains only the latest dataset version. Training MP4s use browser-compatible H.264 High / YUV420P / BT.709 at CRF 1; original full-resolution recordings remain unchanged.

Global meta/stats.json uses training episodes only. meta/episodes/ contains episode boundaries and per-episode descriptive statistics. Video shards are concatenated by stream copy without re-encoding.

Per-camera signal associations

Each 30 Hz row includes camera_signal_associations, a JSON string mapping each camera to its actual image receive timestamp and nearest fresh body observation, VR head event, VR controller event (including buttons), and successful SDK command. Left/right VR controllers are matched independently; vr_controllers additionally records events where both poses are valid. Missing left tracking never discards a real right-controller event. Full matched values and original source-file/line references are embedded. Maximum absolute match interval is 100 ms. camera_<signal>_valid and camera_<signal>_dt_s expose per-camera validity and signed signal-minus-image offsets. VR browser source clocks and camera hardware exposure clocks are not verified as synchronized; matching uses host receive time. Missing signals remain null and never erase real images. Main 31-dimensional state/action labels and their training masks are unchanged. See meta/camera_signal_alignment.json for measured coverage.

Training and replay share the same canonical 30 Hz control knots obtained through 250 → 60 → 30 Hz. Complete control.* columns retain holds even where sample_valid excludes a training example. Reconstruct 250 Hz with the saved per-run PCHIP/linear method in meta/replay_interpolation, using code/canonical30_dataset.py:read_episode; never filter a second time or stretch the original clock. All 30 episodes reproduce the canonical 250 Hz arm and right-gripper sequence bit for bit. Leg, waist and left-gripper action labels remain explicitly measured-trajectory proxies, not published SDK commands.