Morphology-Adaptive Muscle-Driven Locomotion via Attention Mechanisms

This paper introduces an attention-based approach to muscle-driven locomotion that allows a single controller to adapt to different body plans. We consider physically simulated agents with bodies modeled as deformable meshes and controllable fibers that act as muscles. Vertices serve as addressable memory, defining keys and values, while muscles generate multi-head queries to extract relevant information from vertex states and produce actuation signals. A key characteristic of our approach is that all muscle fibers share the same attention-based control policy. This enables locomotion controllers that accommodate any number of inputs and outputs while also remaining permutation invariant. We demonstrate the effectiveness of our method by successfully controlling both bipedal and quadrupedal body plans using the same control policy.