Projects
Learn to Teach
Learn to Teach (L2T) is a one-stage privileged learning framework for the Digit humanoid. Instead of training a privileged teacher and then distilling it afterward, L2T learns teacher and student policies together, reusing simulator samples through shared dynamics to reduce training time and sample complexity. The RL variant transfers zero-shot to hardware and walks across outdoor, indoor, slippery, rocky, grassy, sandy, and disturbed settings.
Distributional Inverse Reinforcement Learning
We introduce a distributional framework for offline inverse reinforcement learning that models uncertainty over reward functions and the full distribution of returns, rather than estimating a single reward. Using first-order stochastic dominance and distortion risk measures, the method captures richer structure in expert behavior, enabling risk-aware imitation and state-of-the-art performance across synthetic, neurobehavioral, and MuJoCo benchmarks.
SEEC: Stable End-Effector Control
A model-enhanced residual learning controller for humanoid loco-manipulation. SEEC pairs a model-based stabilizer for end-effector tracking with a learned residual that handles unmodeled dynamics during locomotion, so the policy only has to correct what the model gets wrong.
RL-augmented Adaptive Model Predictive Control
Adaptive MPC for bipedal locomotion augmented with reinforcement learning: the RL agent corrects the parametric dynamics model online so the controller stays accurate over challenging terrain, while MPC keeps the predictability and constraint handling that pure RL controllers lack.