twanghcmut/g1-dex3-throw-showcase

Unitree G1 + Dex3: pick and throw into a box (Isaac Sim)

A Unitree G1 (pelvis fixed) picks a real household object with its Dex3 hand and throws it into a box 1–2 m away. Clips are recorded episodes, not replays; the intro's last segment shows the failure cases that are filtered out of the dataset.

Waist locked (arms + hands only, 14 joints)

With waist (15 joints, includes countertop picking)

Slow motion (×3)

Side view + the robot's head camera (the head camera is what a VLA policy would see)

More in clips/. The full episodes (joint states, actions, both cameras, instructions) are stored as a LeRobot v2.1 dataset (not published here).

What is randomised

Objects	68 real RoboCasa / RoboTwin meshes (apples, cans, sauce cans, jam jars, Rubik's cubes, tea, coffee, milk and food boxes) + 7 primitives
Physics	mass ×0.7–1.3, centre of mass ±5 mm, object friction 0.4–1.2, restitution 0–0.2, finger friction 0.7–1.3, arm gains ±15 %, finger gains ±20 %
Scene	object pose and yaw on the pedestal, target box position (0.9–2.0 m), box / pedestal / floor / wall colours, two distractor objects
Rendering	dome + sun light intensity, colour and direction; head-camera extrinsics jitter; random third-person viewpoint
Language	6 instruction templates × object name × box colour

Robot limits are kept close to the hardware: arm torques and speeds from Unitree's G1 29-DoF URDF, Dex3-1 finger torque 0.76 Nm (1.5 Nm geared joint) from Unitree's spec sheet.

How episodes are generated

1. Grasps: GraspGenX (unitree_g1 gripper = Dex3) proposes 64 grasps per object once, offline, in the object's frame. At run time the best reachable one is chosen; differential IK (waist + 7-DoF arm) moves the hand there, re-planning to the next grasp if the hand cannot reach the first.
2. Throw: a 12-parameter motion primitive (absolute start and release postures, wrist pitch, swing time, release speed, finger-opening lead) executed open loop as quintic joint trajectories.
3. Search: batched cross-entropy method over the 12 parameters, 256 parallel environments, no rendering, with warm starts from the nearest solved target. ~12 episodes/s, 1,214 successful throws in 17 min.
4. Scoring: each solution is re-thrown with other objects and fresh physics to estimate how robust it is.
5. Collection: fresh randomised episodes reuse nearby solutions, run with cameras, and are kept only if the recorded throw lands in the box.

Known limitations

• Pelvis is fixed (no balance control).
• Thin or round objects (soap bars, cups, computer mice, balls) are excluded: the Dex3 grasp rarely holds them.
• Physics and rendering are simulated; nothing here has been validated on hardware.