research ReSCoT-Q Reference-guided surface-frame control for quadrupeds — force-controlled tool sliding on Unitree Go1 hardware with MPC and whole-body control. GaitNet Learning-based acyclic footstep planning for dynamic quadruped locomotion — CNN foothold cost maps + RL gait selection on top of perceptive MPC. projects Humanoid Ice Skating Emergent ice-skating locomotion for a Unitree G1 via blade-contact RL — no motion capture, no imitation data. Skating falls out of the physics. Bimanual Lifting Human-like bimanual object lifting and carrying for a Unitree G1 via residual AMP–PPO over a frozen locomotion policy. Flow Matching Policies Conditional Flow Matching visuomotor policies in robomimic — beats Diffusion Policy head-to-head, with a 94× faster 1-step inference mode. Acrobatic Quadrotor Flight SE(3) geometric control with L1 adaptive augmentation on a Crazyflie 2.1 — zero-radius flips, recovery, and 94% disturbance rejection in ROS 2 simulation. Go1 PPO Locomotion PPO locomotion for the Unitree Go1 in JAX/MuJoCo MJX — asymmetric actor-critic, domain randomization, rough terrain, and real-robot deployment.
