geonmin-kim/robodojo-v3

RoboDojo bimanual demonstrations (ARX X5)

Scripted-expert demonstrations collected in RoboDojo (Isaac Sim 5.1) on a two-arm ARX X5 cell, exported as a LeRobot v3 dataset.

What is in it

Episodes	145
Frames	23,608
fps	12
Robot	arx_x5_bimanual
Distinct instructions	39

Task	Episodes	Frames	Mean length
general_pickup	102	7,782	76
stack_blocks_by_language	43	15,826	368

Episodes are successful ones only: every episode here reached its task's success milestone, as judged by the benchmark's own reward code.

Observations and actions

Three cameras at 640x480 - observation.images.top, .left_wrist, .right_wrist - plus a 14-dimensional state and action:

[left_x, left_y, left_z, left_roll, left_pitch, left_yaw,
 right_x, right_y, right_z, right_roll, right_pitch, right_yaw,
 left_gripper, right_gripper]

Poses are end-effector poses in the environment frame, in metres and radians. Grippers are normalised, 1.0 open and 0.0 closed.

One thing to know before training on this

observation.state is the commanded pose of the previous control step, not a measured one, and every frame carries state_is_commanded = 1 saying so.

The simulator reports TCP only at decision boundaries, so proprioception between decisions is not recorded anywhere. Rather than drop to one frame per decision - seven frames for an episode that ran seventy-seven control steps, which no action-chunking policy can learn from - this export reconstructs the control-rate timeline from the densified trajectory the host solved IK against, and pairs it with camera frames decoded from the recorded videos.

That is a self-consistent end-effector formulation and is honest as long as tracking is tight. Measured translation error at the decision boundaries of these runs is 0-2 mm when settled. Treat it as commanded-pose supervision, not as measured proprioception.

control_step carries the simulator control step each frame came from.

How the demonstrations were produced

A scripted expert plans from the simulator's ground-truth object poses, one tool call per turn. It is not a learned policy and not a teleoperator: it is the benchmark's own geometry, driven through the same tool interface an agent would use. Grasp poses come from a measured table built by sweeping, per object, three points along the long axis x four yaw angles x three heights and recording which closures survived a clearing rise.

Instructions are rebuilt from each episode's evaluation layout file rather than taken from the run record. The simulator's recorded instruction does not reliably name the object that the layout actually placed - measured over a 156-episode batch, 104 of them named a different object - so episodes whose instruction could not be checked against their layout are excluded rather than shipped with a sentence that may not match the motion.

Known limits

• Two of the benchmark's tasks only. match_and_pick_from_conveyor and plug_in_charger are not included: the scripted expert does not solve them.
• Long tail. The most common single instruction has 19 episodes and many have two or fewer.
• Frame counts are unbalanced against episode counts: stack_blocks_by_language is a third of the episodes but two thirds of the frames, because its episodes are about five times longer.
• Simulation only. No real-robot data, and no force or tactile channel - the simulator exposes none.