zjrwtxtechstudio/soarm101_randomized_mujoco_v1

SO-ARM101 randomized MuJoCo pick-and-lift

Visualize in LeRobot

Generated from the local SO-101 MuJoCo simulation in zjrwtxtechstudio/soarm101_randomized_mujoco_v1. It is intended as a randomized synthetic extension to zjrwtxtechstudio/soarm101__pickdata__01_20261103_190516.

This is a generic object pick-and-lift auxiliary dataset, not the source video's cable-to-groove task. Use zjrwtxtechstudio/soarm101_cable_groove_randomized_v1 for the task-specific extension.

Contents

• Episodes: 48
• Frames: 1456
• Objects: 48
• Successes: 24
• Failures: 24
• FPS: 5
• Cameras: 240x320 RGB

The dataset uses six-dimensional SO-101 joint positions for action and observation.state. It also includes end-effector pose and EEF10 fields. context.split_index is 0/1/2 for train/validation/test object identities.

Randomization

• 48 Google Scanned Objects across containers, bottles, tools, electronics, toys, irregular objects, packages, and rings.
• Object identity, spawn position, yaw, initial joints, scale, mass, friction, servo gain, and the overview camera are randomized.
• Every fourth object episode is attempted with action noise after the grasp approach; physically invalid rollouts are retried without noise.
• The overview camera is jittered around its validated pose; the wrist camera remains attached to the gripper.

Provenance

• Robot: RobotStudio SO-101 MuJoCo model.
• Objects: Google Scanned Objects via kevinzakka/mujoco_scanned_objects, CC-BY-4.0.
• Trajectories: scripted grasp expert, not a learned policy.
• Rendering: MuJoCo OSMesa software rendering.

This is simulation data. Simulator success is not a real-robot success claim.