axiboai/robomind2-ee

RoboMIND 2.0 (selected real-robot parts) with end-effector poses, 224 px video

A modified copy of part of RoboMIND 2.0 (X-Humanoid; arXiv 2512.24653), from its official per-episode release on ModelScope (X-Humanoid/RoboMIND2.0-*). One LeRobot v3.0 dataset per robot, each in its own folder:

Folder	Robot	Episodes	Frames	Source task folders	Instructions	fps label	Video
franka/	franka_fr3_dual	85661	20941113	157	147	13	cam_high 224x126, cam_left_wrist 224x168, cam_right_wrist 224x168
ur5/	ur5e_dual	49929	9490022	143	137	7	cam_high 224x126, cam_left_wrist 224x168, cam_right_wrist 224x168
agilex/	agilex_piper_dual	16655	33284681	56	55	100	cam_high 224x168, cam_left_wrist 224x168, cam_right_wrist 224x168
agilex_arx5/	agilex_arx5_dual	3754	6348675	16	15	63	cam_high 224x168, cam_left_wrist 224x168, cam_right_wrist 224x168
ark/	arx_x5_dual	13248	14438176	54	54	53	cam_high 224x168, cam_left_wrist 224x168, cam_right_wrist 224x168
ark_wide/	arx_x5_dual	181	207558	3	3	53	cam_high 224x126, cam_left_wrist 224x168, cam_right_wrist 224x168

Source task folders: the source's task folders the episodes come from. Instructions: distinct instruction strings (meta/tasks.parquet, total_tasks in meta/info.json); some folders share an instruction.

Changes from the source

• Selection: the two-arm robots on a fixed base with two-finger grippers and two wrist cameras: the source's Franka-Part-1 to -5 (episodes of the six-camera setup only), UR5, Agilex (split by arm type into agilex/, Piper arms, and agilex_arx5/, the older ARX5-type arms) and Ark (ark/). In ark/ the source's camera_left and camera_top keys swap between episodes; each episode's head and left wrist camera were told apart from the images (the wrist camera sees its gripper fingers) and the arm motion. Episodes not included, with the reasons:
  • franka: 26964 of 112625 source episodes (26944: 4-camera rig without wrist cameras; 18: recorded EE pose does not match the kinematics (off by more than 10 mm); 1: corrupt image in the source file; 1: source file is empty (a 6-13 KB stub with no arm or camera data))
  • ur5: 4512 of 54441 source episodes (4501: source file is empty (a 6-13 KB stub with no arm or camera data); 9: leader arm pose does not match the follower arm (off by more than 600 mm); 1: source file is missing camera(s): camera_top, camera_wrist_left, camera_wrist_right; 1: source file is missing camera(s): camera_wrist_right)
  • agilex: 5551 of 22206 source episodes (3988: not a Piper arm (older ARX5-type arm: in agilex_arx5); 698: source file is missing camera(s): camera_left, camera_right; 477: left arm's recorded joints never change (frozen stream); 363: right arm's recorded joints never change (frozen stream); 25: leader arm pose does not match the follower arm (off by more than 600 mm))
  • agilex_arx5: 18452 of 22206 source episodes (16910: not an ARX5-type arm (Piper arm: in agilex); 698: source file is missing camera(s): camera_left, camera_right; 477: left arm's recorded joints never change (frozen stream); 363: right arm's recorded joints never change (frozen stream); 4: recorded EE pose does not match the kinematics (off by more than 10 mm))
  • ark: 1369 of 14617 source episodes (936: left arm's recorded joints never change (frozen stream); 394: head and wrist cameras could not be told apart; 28: right wrist camera frame size differs from the other episodes of its session; 8: different right-wrist camera, too few episodes to keep; 3: left gripper values have no known open/closed scale)
  • ark_wide: 59 of 240 source episodes (40: head and wrist cameras could not be told apart; 19: head camera frame size differs from the other episodes of its session)
• ark_wide/ is the same robot as ark/, from sessions with a different camera: cam_high 224x126 instead of 224x168. A LeRobot dataset has one frame size per camera, so these episodes are a separate folder; everything said here about the robot applies to both.
• observation.state and action: 16 float32 values, [x y z qw qx qy qz gripper] for the left arm, then the right arm. Absolute poses computed by forward kinematics from the source's joint positions: the follower arm's measured joints for the state, the leader arm's joints of the same row for the action.
  • franka: two Franka arms on one table; kinematics: the Panda chain panda_link0 -> panda_link8 (frankaemika franka_description; FR3 has the same link lengths), which reproduces the recorded EE poses. Grasp point: 0.1034 m from the flange (Franka Hand TCP, franka_description panda_hand_tcp).
  • ur5: two UR5e arms on one table; kinematics: Universal Robots' UR5e DH parameters, base -> tool0, which reproduce the recorded EE positions (the recorded EE rotation differs by a few degrees and is not used). Grasp point: 0.15 m from tool0 (the source does not name the two-finger gripper; the value of a Robotiq 2F-85-class gripper).
  • agilex: AgileX Cobot Magic (parked), Piper arms; kinematics: the official Piper URDF chain base_link -> link6 (agilexrobotics/piper_ros); the source's EE fields are zeros. Grasp point: 0.1358 m along link6 z (the finger joints of the official URDF).
  • agilex_arx5: AgileX Cobot Magic (parked), the older ARX5-type arms of the same repo; kinematics: the ARX X5 URDF base_link -> link6 (the X5A model), which reproduces the source's recorded EE positions and roll/pitch/yaw. Grasp point: 0.12 m along link6 x (the ARX X5 gripper).
  • ark: ARX dual-arm station (parked, the Ark repo); kinematics: the ARX X5 URDF base_link -> link6, which reproduces the source's recorded EE positions. Grasp point: 0.12 m along link6 x (the ARX X5 gripper).
  • ark_wide: ARX dual-arm station (parked, the Ark repo); kinematics: the ARX X5 URDF base_link -> link6, which reproduces the source's recorded EE positions. Grasp point: 0.12 m along link6 x (the ARX X5 gripper).
• Pose frame: x = the gripper's approach axis, y = the axis the fingers close along, z completes the right-handed frame; origin 12 cm behind the grasp point. World frame: 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), metres, with the arm's forward direction along +y. Quaternions are w-first and continuous inside an episode.
• Gripper: 0 closed .. 1 open (the source's values rescaled and clipped; in ark/ each arm's scale is read from its own values). In ark/ and agilex_arx5/ the action's gripper, and in ark/ the whole action, is the follower arm's value of the next frame (the source's leader stream does not match the follower there).
• Video: three colour cameras per robot, cam_high (top or front view), cam_left_wrist, cam_right_wrist, resized to 224 px on the long side (aspect kept) and encoded as AV1 (crf 30), every source frame kept. Depth, the other cameras and the source's other fields are not included.
• Timing: the source records at a rate that differs between episodes; every frame is kept (no resampling) and timestamps are frame number / the fps label of the folder.
• Language: one instruction per episode, the source's language_instruction.

License and attribution

RoboMIND 2.0 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 (X-Humanoid); please cite RoboMIND 2.0 (arXiv:2512.24653) when you use it.