Architecture Is All You Need: Diversity-Enabled Sweet Spots for Robust Humanoid Locomotion
This work addresses the problem of enabling robust locomotion for humanoid robots in real-world scenarios, representing an incremental improvement through architectural innovation rather than a new paradigm.
The paper tackles robust humanoid locomotion in unstructured environments by proposing a layered control architecture that separates fast proprioceptive stabilization from slower perceptual decision-making, achieving substantially more robust performance than monolithic end-to-end designs, as demonstrated on a Unitree G1 humanoid across stair and ledge tasks where one-stage policies fail.
Robust humanoid locomotion in unstructured environments requires architectures that balance fast low-level stabilization with slower perceptual decision-making. We show that a simple layered control architecture (LCA), a proprioceptive stabilizer running at high rate, coupled with a compact low-rate perceptual policy, enables substantially more robust performance than monolithic end-to-end designs, even when using minimal perception encoders. Through a two-stage training curriculum (blind stabilizer pretraining followed by perceptual fine-tuning), we demonstrate that layered policies consistently outperform one-stage alternatives in both simulation and hardware. On a Unitree G1 humanoid, our approach succeeds across stair and ledge tasks where one-stage perceptual policies fail. These results highlight that architectural separation of timescales, rather than network scale or complexity, is the key enabler for robust perception-conditioned locomotion.