Jie Xu

Robot imitation learning does not require cumbersome active tracking suits: a single monocular camera observing passive wrist markers and sparse workspace fiducials can reconstruct metric 6-DoF trajectories offline.

Worst-case adversarial bandits no longer have to learn from scratch: transferring predictable task-level priors across random action sets achieves provably sublinear transfer regret with intrinsic-dimension $\mathcal{O}(\sqrt{n})$ guarantees.