Xianyuan Zhan

Supervising first-order action dynamics alongside actions themselves unlocks a dramatically more robust and performant policy class for complex control tasks.

ODEWorld achieves high-quality long-horizon predictions and planning-oriented dynamics by seamlessly integrating continuous-time modeling with ODE-based representations.

Diffusion models can achieve a 10x performance boost in real-world autonomous driving when optimized for trajectory representation, loss space, and augmented with reinforcement learning.