I was the tech lead of a team working on manipulation with a Unitree G1 humanoid and NVIDIA GR00T policies. My own work was on simulation and reinforcement learning. I fine-tuned a frozen GR00T policy with Diffusion Steering RL, built an RL pick task with a dexterous hand, and wrote a VR teleoperation system.
I led the technical work of the team from May to November 2026 and was responsible for simulation and reinforcement learning. I also wrote the team's documentation for policy inference and teleoperation.
The GR00T N1.7 policy generates actions with a diffusion model. In Diffusion Steering RL the policy weights stay frozen and a small RL agent learns to choose the noise that the diffusion model starts from. I applied this to a cube-sorting task with the G1 in MuJoCo. Full-task success (all three cubes sorted) increased from 48.3% to 78.3%. Success per cube increased from 78.9% to 88.9%.
I built a pick task for the G1 with an 11-DOF BrainCo Revo2 hand in MuJoCo-Warp and trained it with PPO on 8 A100 GPUs. The first policy reached 96% success. When I looked at the videos I found that the hand started inside the cube, so the policy did not really grasp. I corrected the initial state and changed the success metric so that it could not be satisfied by throwing the cube. With the corrected task the policy reached 88 to 90% over three seeds.
I wrote a teleoperation system for the G1 arms with a Meta Quest 3. It uses WebXR, inverse kinematics solved as a QP, and stereo video sent back to the headset. Six different dexterous hands can be attached to the same arms (Inspire, LEAP, Allegro, Shadow, ORCA and BrainCo Revo2), and only the finger mapping changes. Episodes are started and stopped with hand gestures.
I also built a simulation of a factory workstation from reference videos for testing packing tasks.