adityx23/Pi-Atlas_Sanity_Check_Dataset

YOR robot — LeRobot Dataset v2.1

399 episodes, 270453 frames at 30 Hz, 9 task prompts:

• Place the Red Chili on the Plate using the right arm: 100 episodes
• Place the Red Chili in the Basket using the right arm: 50 episodes
• Place the Eggplant on the Plate using the right arm: 50 episodes
• pick up the eggplant: 34 episodes
• go to the basket in the workshop: 33 episodes
• place the eggplant in the basket: 33 episodes
• pick up the eggplant from the basket: 33 episodes
• go to the table in the living room: 33 episodes
• place the eggplant on the table in the living room: 33 episodes

Offline derivative of exports/pi_atlas_manifest_20261027.json. Source recordings remain unchanged. Conversion configuration and per-episode provenance are in meta/conversion/. Recorded success/failure labels are retained; no success adjudication is implied. Depth is not converted. RGB is resized to 640x360, aspect ratio 16:9 by default.

Control contract

observation.state is 21D: left joints(7), right joints(7), measured left/right grippers(2), measured neck yaw/pitch(2), measured base vx/vy/wz(3). action is 21D: commanded left joints(7), commanded right joints(7), commanded right gripper(1), commanded base vx/vy/wz(3), lift direction(1), commanded neck yaw/pitch(2). Joints are radians, grippers 0=closed/1=open, base m/s and rad/s, neck degrees, lift direction -1/0/1. This is a custom YOR schema, not ALOHA. The left gripper is deliberately excluded from the main action; deployment must leave/control it separately. The source lift measurement is not in the main state because it is missing in early episodes. The lift command is genuinely recorded, including zero.

The original 17D measured state is auxiliary.observation_state_full with component mask annotation.observation_state_full_valid. Missing components are represented by zero ONLY in auxiliary data and always have a false mask. The sparse left-gripper command is auxiliary.left_gripper_command with annotation.left_gripper_command_valid. These auxiliary sentinels must not be treated as observed zero targets.

Alignment and training masks

Each episode is resampled on a fixed 30 Hz grid within its robot-data interval. Robot vectors are nearest Orin recv_ns; cameras are nearest Orin capture_ns. Pi robot timestamps are retained for provenance, not used for synchronization. Camera stale/missing frames are black, with annotation.camera_valid.=false. Robot rows farther than 100.0 ms remain nearest observed values but annotation.robot_valid=false: this avoids inventing an interpolation. Camera threshold is 100.0 ms. Source matches and offsets are retained per row.

annotation.frame_valid means a fresh robot row AND a fresh ZED-left frame; it does not require wrist availability, task success, or IK success. Use wrist validity as model image masks. annotation.right_ik_ok and annotation.right_ik_known permit an additional filter. Standard LeRobot/openpi training does NOT automatically apply these custom masks. A custom adapter/data sampler must consume them and prevent action chunks crossing invalid spans. meta/conversion/valid_windows.jsonl records contiguous valid intervals and valid chunk-start ranges for 50-frame windows. Ranges are half-open. Use observed_only_valid_runs to exclude inferred right-gripper labels. Inferred prefixes do not reconstruct limp/torque-enable state.

annotation.left_arm_command_valid is false where the recording had no left-arm command (right-arm-only tasks never command the left arm); there the left-joint action equals the measured left joints, i.e. hold still. The seven verified right-gripper prefix overlays (eggplant_to_plate_76 only) are applied when the repair manifest is enabled. annotation.right_gripper_inferred flags their resampled rows. Apart from the flagged left-arm hold rows, no recorded joint commands are replaced with measured joints. Source raw JSON and repair provenance remain in the raw archive.

All episodes appear in the default train split only as a storage convention. Build an episode/session/day validation split and recompute normalization on the actual training split before fitting a policy. Episode statistics describe stored resampled values, including black image placeholders and nearest samples at invalid timesteps; do not use them blindly for a masked/filtered training subset. A format conversion is not a declaration that every frame is ready to train.

Format reference: https://github.com/huggingface/lerobot/tree/v0.3.3/src/lerobot/datasets (CODEBASE_VERSION v2.1). Current LeRobot releases may use v3. Use a compatible v2.1 loader or explicitly port this output.