Reinforcement-learned locomotion policies are typically shallow networks trained in simulation, but how they internally represent motion is poorly understood. This paper proposes looking at the effective rank of the network Jacobian, but with a twist: instead of a single global value, the rank is conditioned on the phase of the gait cycle. That phase-resolved view reveals structures that a global measurement hides.
The authors find that architectural choices—specifically observation and layer normalization combined with residual connections—allocate more than one additional dimension of effective rank to swing-dominant states compared to stance-dominant states. Vanilla MLPs mostly lack this pattern. This suggests that a policy's capacity is not spread uniformly across the gait cycle, and that the distribution is tied to architectural design.
Building on this, the authors turn the phase-conditioned rank into a practical diagnostic for policy health. Policies that score well on this diagnostic exhibit up to 2x lower joint jitter, both in simulation and on a physical Spot robot. The work links representation structure to sim-to-real transfer, offering a new way to evaluate and design locomotion policies beyond task performance alone.