G1 Locomotion and Whole-Body Learning
Simulation studies of humanoid walking, supported pickup, and large-object enclosure.
From locomotion to contact-rich tasks
In Isaac Lab, I have studied Unitree G1 locomotion and manipulation tasks that need the body as well as the hands. A walking policy passed a frozen simulation evaluation. The pickup and large-object enclosure studies exposed harder control and validation problems: assisted pickup and relaxed enclosure were observed, while unassisted pickup and sustained strict enclosure remain open.
These results are simulation-only. They do not establish transfer to a physical G1 or safe autonomous behavior. The current work uses them to define better experiments for learning contact-aware whole-body skills.
Simulation replays
G1 in VR: Human World prototype
A separate Quest/G1 control prototype places the headset at a room-scale human viewpoint independent of the simulated robot’s body motion. Both hands specify world-space palm targets; bounded inverse kinematics and a posture controller generate the robot motion. The simulation videos above show different Isaac Lab policies, not this Quest controller.
The Human World implementation passed 39 targeted CPU tests. Five standing validation episodes in Isaac simulation had no recorded falls or joint-limit violations, with reachable palm-target 95th-percentile error at most 1.67 cm. A walking diagnostic stayed upright but missed its tracking target, so walking remains experimental. The Quest app was installed, but a GPU conflict blocked the live Human World session. On-device acceptance and physical G1 operation remain unverified.