Junting Dong

Ditching brittle hand-estimator-plus-SLAM cascades for an end-to-end streaming memory network slashes world-space hand tracking error by 21.4% while doubling inference throughput.

Robots can now catch dynamically moving objects with human-level dexterity, thanks to a shared autonomy framework that intelligently blends teleoperation with learned diffusion policies.

Monocular SLAM gets a serious upgrade: M^3 slashes pose estimation error by 64% by fusing multi-view foundation models with dense matching.