Humanoid running and jumping need more than motion: Contact-UAN patches the actuator force model with 6.5 minutes of walking data

When humanoid robots move from simulation to hardware, the motion can look right while the force interactions learned in training do not come along, showing up as poor torque efficiency, weak load regulation and fragile impact response. Earlier residual actuator learning relied on replaying hardware trajectories to supply the training signal, and if replay began from a simulator state not equivalent to the hardware, the learned correction absorbed reset artifacts instead of actuator discrepancies. On the hard target simulator, downstream policies trained with the contact-consistent-reset UAN raised 20-second survival from 49.8% with direct-estimator resets to 96.7%, close to the 96.6%–98.4% of component or full-oracle resets.

Contact-UAN corrects only the command the simulator receives, leaving rigid-body dynamics, the contact solver and numerical integration untouched, so the simulator's physics priors are preserved. A fully learned world model must fit all robot-environment dynamics from data, while supervised actuator networks need reliable output-torque sensing that many humanoid platforms lack. On held-out running hardware data, replaying identical recorded commands from the same relabeled initial states, Contact-UAN cut the all-joint aggregate torque mismatch by 69.2%; on the hard target simulator, even with ground-truth initialization the residual reduced 24-step joint-position RMSE from 0.0720 to 0.0184 rad.

The contact-consistent reset state is inferred from proprioception, including the simulator-side stance-foot pose and base velocity, so replay begins with contact forces consistent with the hardware. Matching a suspended free-space motion in simulation is only a kinematics problem and lacks the contact information needed for stance, force generation, impact and landing, while millimeter-scale foot-height and degree-scale orientation errors substantially change simulated contact forces. The state estimators are trained in simulation and regress pose only on confident stance samples; on the hard target simulator, the contact-conditioned estimator reduced contact-height RMSE by an order of magnitude.

One UAN trained on 6.5 minutes of real walking data is reused across tasks, covering running, force and impedance control, and jumping. The no-UAN baseline triggered motor-overheating protection and entered safe-damping mode at a high-speed command, had larger force-tracking error, and its task-reward-fine-tuned jumping policy failed during landing. Running used five hardware trials per command at three commanded speeds, with the UAN policy holding velocity tracking while using less torque than the baseline; force tracking used a dynamometer with four trials per commanded force and cut overall error by 43%; jumping was fine-tuned with task reward after reference tracking, and the UAN policy transferred with stable landings.

The demonstrated scope is a 29-DoF Unitree G1 humanoid with a nominal simulator using BeyondMimic G1 parameters and about 6.5 minutes of real walking data; the downstream tasks are running, force and impedance control, and jumping, all validated on hardware. The state estimators assume flat terrain and are used only for offline relabeling, not in downstream policy training or deployment.

The current model does not capture coupled actuator effects or effects that depend on motor temperature; temperature effects are hard to identify from short datasets because motor temperature evolves slowly relative to the control loop. In sim-to-sim, adding task-specific running data gives little additional benefit over walking data alone. Alternative architectures show similar downstream performance in a matched simulation comparison.