brandonyang/artem-fold-towel-filtered
Artem fold-towel filtered trajectories Observation-only LeRobot v3 derivative of brandonyang/artem-fold-towel. It contains 781 demonstrations (1048134 frames) accepted by the continuous bimanual YAM replayability pipeline. The 14-D observation.state contains the smoothed, trajectory-optimized YAM-achievable UMI1 pose, normalized UMI1 gripper, UMI2 pose, and normalized UMI2 gripper. The two original UMI videos, timestamps, frame cadence, and task are preserved; action is… See the full description on the dataset page: https://huggingface.co/datasets/brandonyang/artem-fold-towel-filtered.
Artem fold-towel filtered trajectories
Observation-only LeRobot v3 derivative of brandonyang/artem-fold-towel. It contains 781 demonstrations (1048134 frames) accepted by the continuous bimanual YAM replayability pipeline.
The 14-D observation.state contains the smoothed, trajectory-optimized YAM-achievable UMI1 pose, normalized UMI1 gripper, UMI2 pose, and normalized UMI2 gripper. The two original UMI videos, timestamps, frame cadence, and task are preserved; action is intentionally absent.
The source dataset already stores each gripper in [0, 1]; those values are preserved and clipped to the same range. The formula, observed source extrema, and clipping counts are recorded in meta/filtering.json.
IK and FK use the current BiYAM-pinned I2RT YAM model with the custom 220 mm linear gripper and UMI-compatible grasp frame.
Initial-state convention: optimized poses use the shared frame anchored to the measured initial UMI1 pose; no optimized start state is forcibly zeroed, and UMI2 preserves the measured bimanual relative geometry. See meta/filtering.json for pinned source revisions, gates, provenance, and the output-to-source episode map. Collision checks are sampled-capsule proxies; validate the exact scene before hardware playback.