Xiang Yin

Offline reinforcement learning can achieve zero-shot generalization to unseen Signal Temporal Logic tasks by decoupling reachability learning from task-conditioned planning.

Generative image restoration has flipped from struggling to add detail to now struggling to control the quality and semantics of the details it hallucinates.

Skip the hand-engineered maps: this end-to-end framework learns to plan robot trajectories directly from camera feeds and temporal logic constraints.