axiboai/abc130k-ee

ABC-130K with end-effector poses

A modified copy of ABC-130K (Allshire et al., "Scalable Behavior Cloning with Open Data, Training, and Evaluation", arXiv:2606.27375), in its LeRobot v3.0 release lerobot/abc_130k_v3_train and lerobot/abc_130k_v3_val. The source records joint positions; this copy stores end-effector poses. LeRobot v3.0, 130811 episodes, 387528028 frames, 201 instructions, 30 fps.

Changes from the source

• observation.state and action: 16 float32 values instead of the 14 joint values, [x y z qw qx qy qz gripper] for the left arm, then the right arm. Absolute poses, computed by forward kinematics of the I2RT YAM arm (the chain of the open ABC station model, github.com/amazon-far/abc) from the source's measured joints (state) and commanded joints (action) of the same row.
• Pose frame: x = the gripper's approach axis, y = the axis the fingers close along, z completes the right-handed frame (up when the gripper points forward); origin 12 cm behind the grasp point (0.1347 m along the approach axis from the gripper mount).
• World frame: z up, the station faces +y, the left arm is at -x. Each arm's pose is relative to its own base, which is placed at (-0.3, -0.45, 0.765) (left) and (0.3, -0.45, 0.765) (right), in metres.
• Quaternions are w-first, continuous inside an episode (first frame w >= 0).
• Gripper: 0 closed .. 1 open, clipped to that range (measured in the state, commanded in the action).
• Video files: unchanged (the same bytes as the source). Camera keys renamed: top -> cam_high, left_wrist -> cam_left_wrist, right_wrist -> cam_right_wrist.
• Instruction: the source's task string.
• The splits of the source are in one dataset (meta/info.json splits). Columns other than the standard LeRobot ones are dropped. Per-episode and global statistics of observation.state and action are recomputed.
• 6 source episodes are not included (6: the commanded pose is more than 60 cm from the measured one). Rules: a non-finite value; a measured joint more than 0.5 rad outside the arm model's limits; a commanded pose more than 60 cm from the measured one; rows that do not match the episode's length; a missing camera; a source video that does not cover the episode's rows. Episodes are renumbered without gaps, in source order.

Split	Source episodes	Episodes here	Episode index range
train	129225	129219	0:129219
val	1592	1592	129219:130811

Known damaged video files

Three video files are damaged in the upstream release (lerobot/abc_130k_v3_train, the same bytes) and in this copy: they cannot be decoded past the point given. The episodes stored in them have no usable video for that camera from that point on.

File	Camera	Decodes up to	Episodes affected
videos/observation.images.cam_left_wrist/chunk-007/file-124.mp4	left wrist	frame 21835 of the file	85448 (from its frame 36), 85449
videos/observation.images.cam_right_wrist/chunk-007/file-574.mp4	right wrist	frame 24 of the file	89561-89568
videos/observation.images.cam_right_wrist/chunk-009/file-616.mp4	right wrist	frame 28 of the file	105883-105890

License and attribution

The source dataset is licensed under the Apache License, Version 2.0, and so is this copy (see LICENSE). This copy was changed from the source as listed above. Credit for the data goes to the source's authors; please cite them: Allshire et al., Scalable Behavior Cloning with Open Data, Training, and Evaluation, arXiv 2606.27375.