CrimsonRobot/crimson-first-dateset

Crimson First Dataset

Real-robot right-arm teleoperation demonstrations, exported with the official LeRobot 0.6.1 dataset writer for subsequent X-VLA adaptation.

Selection

Requested raw source episodes: 4 through 55 inclusive from pilot_right_20260924. Raw episode 10 was labeled discarded; raw episode 19 was labeled failure. Neither is included or relabeled. The remaining 50 successful episodes, 7,860 aligned rows at 10 Hz, passed the converter's timestamp/freshness, finite-vector, recorder-loss and identity checks. Export episode indices are consecutive 0–49, not original raw episode numbers. genome_conversion.json preserves source order and per-episode validation reports.

Features

• observation.state: 8 float32 values: right J1–J7 measured positions (radians), then measured right gripper servo position (radians).
• action: 8 float32 values: right J1–J7 absolute runtime commanded positions (radians), then the gripper command actually applied (0=open, 1=close).
• observation.images.top: RGB video from cam_high.
• observation.images.right_wrist: RGB video from cam_wrist.
• Both images are 640×480. The third raw camera is not part of this export.
• Task text is inherited from session metadata, with LeRobot task indices.

Alignment selects the most recent source sample at or before each 10 Hz row. Actions represent same-row runtime commands, not next measured joint positions. Multi-turn joint angles are preserved. Idle periods are not automatically trimmed. Raw camera recordings were about 15 Hz; exported videos are aligned to 10 Hz.

Intended use and limitations

Use X-VLA's configurable action representation (action_mode=auto) with an 8-dimensional real action vector, not a fixed 14-dimensional joint mode. Preserve the documented joint order, units and camera mapping in deployment. Measured gripper state and binary gripper action intentionally have different units.

These are right-arm demonstrations, not bimanual training samples. Validation of file structure and alignment does not establish task diversity, generalization, physical accuracy, model training convergence or safe autonomous deployment. No X-VLA model forward pass or training result is claimed.

Keep the Hugging Face repository private as requested by the data owner. No redistribution license is assigned here; obtain permission from the owner. No train/validation split has been imposed. Split by complete episodes rather than random individual frames to avoid leakage.