Policy-space Interpolation for Physics-based Characters 61

Controllers for physics-based humanoids often rely on reference animations to synthesize plausible motion trajectories for task-driven control. When the controller is a deep reinforcement learning policy, the output of several controllers can be combined to enhance controller robustness and synthesize new motion variations. However, this requires evaluating several policies at each timestep and combining their outputs, which can be computationally costly. In this work, we propose an alternative approach that combines individual controllers using linear interpolation of network parameters, thereby requiring only a single policy evaluation per timestep. Our method employs a graph-based weight regularization strategy to ensure that similar motions generate similar policy weights during training. We show that this technique produces visually indistinguishable outcomes compared to blending controller outputs, and that the approach easily integrates new control policies without retraining existing ones. We further demonstrate that interpolating or perturbing individual layers results in novel variations of the internal motion pattern that cannot be easily achieved by operating on the actions. This opens a path toward improved variability in controller networks by manipulating their weights. Several compelling use cases demonstrate the benefits of our approach, including interactive control and synthesizing motion variations.