Architecture Is All You Need: Diversity-Enabled Sweet Spots for Robust Humanoid Locomotion

TL;DR

The paper argues that robust humanoid locomotion in unstructured environments emerges from architectural design rather than sheer network scale, positing a minimal layered control architecture (LCA) with a fast proprioceptive stabilizer and a slow perception-driven navigator. A two-stage curriculum trains stabilization first without perception, then augments it with a compact local heightmap encoder to enable longer-horizon planning while preserving fast reflex-like control. Through simulation and hardware on a Unitree G1, the authors show that layered, multi-rate designs outperform monolithic end-to-end policies, particularly under distribution shifts, supporting the claim that timescale separation is the key enabler for perception-conditioned locomotion. The work highlights Diversity-enabled Sweet Spots (DeSS) where the combined stack surpasses any single component tuned in isolation, and demonstrates that simple perception filtering and narrow interfaces suffice for robust performance in complex terrains. The practical impact is a computationally efficient, robust framework for bipedal walking that can be deployed with minimal perception and network complexity, broadening applicability in real-world humanoid locomotion.

Abstract

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.

Figures (6)

• Figure 1: A humanoid robot trained to traverse complex terrain through use of a combination perception information and fast proprioception information. Using this input effectively requires the use of structured architecture in order to produce performant and robust results.
• Figure 2: Training and Deployment Overview: both actor and critic are two-stage architectures each with their own perception encoder. The actor receives noisy heightmap information, while the critic receives perfect information, and each receive proprioception history. During deployment, a depth image is filtered and passed through the trained encoder, and the actor combines this with the proprioception history to determine action.
• Figure 3: Layered verses monolithic architectures: while the network architecture may be identical, training in two phases allows them to assume the layered control structure.
• Figure 4: Training rewards from each of the policies. We see that in training, all the policies perform largely identically; we believe that the small deviations may be a function of the network architecture of that component, such as the LSTM's ability to store a hidden state or the CNN's ability to reason more spatially, but may also simply be products of randomness.
• Figure 5: The four hardware experiment tasks. The first two, stair ascent and stair descent, emphasize navigation, while the second two, hinged and soft ledge, emphasize control.