Ruiyu Gou 苟 瑞宇

We introduce Control Operators, a way of encoding arbitrary inputs to neural networks, making building ML-based animation controllers accessible, even for non-technical users. ^{动画人不骗动画人，好训好用 :P}

Can we control a humanoid to perform a stepping stone task with policy only queried at 10hz? This thesis explores the idea of temporal action chunking, an state-and-time-indexed reformulation of reinforcement learning policies.

We learn a factored form of model predictive control to enable controlling complex contact-guided motions interacting with the environment.

We show that a pure reward-driven reinforcement learning approach with a strong-to-weak curriculum can learn to get up naturally from a wide range of fallen states, without motion capture data.