Contact-UAN: Unsupervised Actuator Networks as a Residual World Model for Humanoid Control

Contact-UAN is a residual world model that aligns simulated and real-world humanoid motion by correcting actuator dynamics discrepancies. Using only onboard proprioception, it infers contact forces and initializes replay with consistent contact wrenches, enabling seamless policy transfer. The method achieves 20% lower running torque, 43% lower force-tracking error, and improved dynamic robustness in real-world tasks.