Ce-Wu Lu

Real-world robot policies can learn autonomously from sparse rewards by treating value predictions as task-progress memory, outperforming offline training by $32.8\%$ success rate.

Learning to predict the visual consequences of actions significantly boosts policy performance and generalization, while simultaneously improving world model accuracy.

Results show that robust visuomotor learning benefits from structuring, rather than removing, visual embodiment information, and embodiment canonicalization in 3D point clouds substantially improves human-to-robot policy transfer without robot demonstrations.

LLMs can now plan complex, long-horizon household tasks from ambiguous instructions with significantly improved performance, thanks to a novel reinforcement learning approach that corrects intermediate reasoning steps.

Achieve one-shot skill transfer in imitation learning by disentangling intent from execution, allowing robots to adapt to new environments simply by injecting the "Intent" token from a demonstration.