robosim-so101-pickup-v3

MuJoCo-simulated "pick up cube" training data for the SO-101 robot arm, generated by RoboSim.

How This Data Was Generated

Each episode runs real MuJoCo physics simulation (mujoco-js WASM, MuJoCo 3.3.8):

1. The SO-101 arm MJCF model is loaded with a cube at a randomized position with varied size/mass/friction/color
2. FK-based inverse kinematics computes joint waypoints for each cube configuration
3. Hybrid collision: arm ghosts through cube during approach, then real contact enabled during grasp/lift
4. Position-controlled actuators (kp=100) drive the arm through approach, grasp, lift phases
5. Grasp is simulated via rigid qpos attachment (cube tracks gripper once jaws close)
6. Joint angles are read back from the physics state (capturing dynamics and contact forces)
7. Observations recorded at 30fps in LeRobot v3.0 format
8. Camera images rendered at 10fps via headless Chromium (Three.js URDF scene)
9. Two cameras: cam_high (overhead 640×480) and cam_wrist (gripper-mounted 640×480)
10. Each physics episode generates 1 augmented copy (Gaussian noise + spatial jitter, no video)

Dataset Details

Property	Value
Robot Type	arm
Robot ID	so-101
FPS	30
Physics Episodes	50
Augmented Episodes	50
Total Episodes	100
Total Frames	23316
Physics Success Rate	50/50 (cube lifted > 5cm)
Task	pick_object
Simulation	MuJoCo 3.3.8 (mujoco-js WASM)
Timestep	0.002s (500Hz physics, 30Hz recording)
Solver	Newton, 100 iterations
Format	LeRobot v3.0 (real Apache Parquet)

Joint Channels

6-DOF observations and actions: [shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper]

• Joints 0-4: degrees (matching real SO-101 STS3215 servos)
• Joint 5 (gripper): 0-100% (0=closed, 100=fully open)

Additional feature: observation.velocity — joint velocities from MuJoCo dynamics

Pickup Phases Per Episode

1. Home (0.5s) — arm at rest position, gripper open
2. Pre-grasp (~1.5s) — arm moves above the cube (collisions OFF)
3. Approach (~1.5s) — arm descends to grasp position (collisions OFF)
4. Grasp (~0.8s) — gripper closes, collisions ON (real contact forces)
5. Lift (~3.0s) — arm lifts the cube to 10cm above table (collisions ON)

Phase durations vary per episode due to speed randomization (0.8x-1.2x).

Domain Randomization

Parameter	Range
Cube X position	12-24cm
Cube Y position	-8cm to +8cm
Cube half-size	1.5-2.5cm (3-5cm full)
Cube mass	30-80g
Cube friction	0.8-1.2
Cube color	Random RGB
Robot speed	0.8x-1.2x

50 unique configurations (seeded RNG, seed=42 for reproducibility).

Trajectory Augmentation

Each physics episode generates 1 augmented copy with:

• Gaussian action noise (std=1.5 degrees on joints 0-4)
• Spatial jitter (std=0.5 degree constant offset per joint)
• Seeded RNG (seed=123) for reproducibility

Camera Images

Property	cam_high	cam_wrist
Type	Overhead fixed	Gripper-mounted
Resolution	640×480	640×480
FPS	10	10
Encoding	H.264, CRF 23	H.264, CRF 23
Coverage	Physics episodes only	Physics episodes only

Camera images are rendered via headless Chromium using the Three.js URDF-based scene. Augmented episodes (noise/jitter copies) have joint data only, no video.

Physics Details

• Hybrid collision: arm starts with contype=2 (ghost), switches to contype=3 (real contact) before grasp
• Position actuators: kp=100 with damping=5, armature=0.5 for stable tracking
• Grasp simulation: cube freejoint qpos tracks gripper_link + offset each timestep
• Contact: real arm-cube contact forces during grasp/lift; cube always collides with table

Usage with LeRobot

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("houmans/robosim-so101-pickup-v3")
print(f"Episodes: {dataset.num_episodes}, Frames: {dataset.num_frames}")

Generated by

RoboSim — Browser-based robotics simulator with MuJoCo physics.

License

Apache 2.0