kamada-riku/cup_stacking_simv1

OpenArm cup stacking simulation v1

A MuJoCo simulation dataset exported with the official LeRobot v3.0 writer.

Task

Start with two inverted blue cups arranged front to back on the robot's left and one cup on its right. Use the left arm to place the two left cups side by side in front of the robot, then use the right arm to stack the right cup on top to form a 2-1 pyramid. Release all cups and retreat.

Contents

• 90 successful scripted teacher episodes, 111,876 frames, 30 FPS.
• OpenArm V1 with two arms and grippers.
• Three RGB video streams: torso-mounted front camera, left wrist, and right wrist; 640 × 480, H.264.
• observation.state and action: 16 float32 values, right arm first, then left arm; seven arm joints and one gripper value per side.
• Arm joint positions and position targets are in radians. Gripper values are in meters per finger.
• Stored actions are absolute simulator targets. Relative arm targets and B-spline teachers are derived separately for BSP training; they are not stored in this LeRobot export.
• All included episodes passed the simulator's physical task checks. Failed generation trials are excluded.

Generation and validation

The scripted teacher uses gradual quintic gripper closing and opening. Initial cup positions vary across episodes. This is a limited scripted distribution, without learned recovery demonstrations.

All source frames were decoded and written. The exported dataset was reopened with the official LeRobot reader; action and state values were checked exactly at the beginning, middle, and end of every episode (270 readback samples), along with all three camera streams. See verification.json and simulation_provenance.json for checks, scene parameters, source hashes, and per-episode provenance.

Simulation-to-real status

Scene geometry, cup dynamics, and cameras are provisional and have not been physically calibrated. Simulator gripper units differ from the real recording interface; mapping to real hardware requires calibration. These are synthetic demonstrations, not real teleoperation recordings.