Modeling and Understanding Neural Weights of Humanoid Policies with Diffusion

This research investigates whether high-performing humanoid control networks can be generatively modeled directly through their weights using a latent diffusion framework. The approach combines a linear autoencoder with a Diff, trained on networks from reinforcement learning for the Unitree G1 robot, generates deployable networks capable of real-world motions. The study also identifies a structural principle in the first-layer weights that reflects kinematic priors, observed across both reinforcement learning and imitation learning training.