CoRe-MoE enables humanoid gait across terrains

A Unitree G1 just proved a two-stage AI approach can walk stairs and run over rough terrain.

The work, CoRe-MoE, tackles a long-standing engineering question: how to split gait generation from terrain adaptation without letting them interfere with one another. In many prior efforts, a single policy has to learn both the mechanics of walking and the dynamics of varying surfaces at once, which creates gradient interference and a shifting distribution when terrain appearance changes. Naive mixture-of-experts tricks helped in principle, but training often leaves you with loosely specialized “experts” that don’t reliably cooperate. The researchers propose a clean two-stage reinforcement learning framework to fix that.

In the first stage, the team trains a stable locomotion policy that supports natural walking and running with smooth transitions between the two modes. The aim is a base gait that remains robust across common robot motions, so the robot doesn’t stumble when surfaces tilt or steps appear out of the blue. Only after that base is established do they introduce a terrain-aware MoE branch. This second stage uses a contrastive objective to shape the gating network so it forms structured terrain representations and prompts the experts to specialize more clearly in different environmental contexts. The final action is a weighted fusion of the base gait output and the terrain-aware branch, which preserves the stability of the underlying motion while letting the robot adjust to terrain idiosyncrasies.

Tests in simulation claim improvements in success rate, locomotion stability, and multi-terrain adaptability over baseline approaches. More telling for practitioners, the authors report a successful zero-shot deployment on a Unitree G1 humanoid, where the robot demonstrated robust walking and running across stairs, slopes, steps, obstacles, and unstructured outdoor terrains. In other words, the policy learned in simulation translated to real hardware without per-terrain retuning, a notable milestone for field-ready humanoids.

For engineers scanning this advance, several concrete implications stand out. First, decoupling gait generation from terrain adaptation can reduce gradient interference enough to allow cleaner specialization among experts. In practice, that means you’re less likely to see a single policy wrestle with competing objectives on a new surface, which translates to more predictable transitions between walking and running. Second, the contrastive objective used to train the terrain-aware gating network helps the system form discriminable terrain representations, making expert selection more reliable when the robot encounters unfamiliar ground. Third, fusing a stable base gait with a terrain-responsive branch provides a safety margin: if the terrain branch misreads a surface, the base gait keeps the motion grounded and smooth.

Yet the approach is not a silver bullet. Real-time inference on embedded hardware remains a constraint, particularly when the perception stack must supply timely terrain cues for gating decisions. There is also the question of how the gating network will behave on truly unseen terrains or under sudden dynamic disturbances, and what failure modes look like when perception latency or sensor noise creeps in. The industry will want to see broader validation across more robot platforms, longer-duration field trials, and a careful accounting of energy expenditure during terrain switching.

Still, the CoRe-MoE result is a meaningful step toward practical, robust legged autonomy. It reframes how engineers think about multi-terrain locomotion: not by stuffing more complexity into a single policy, but by organizing expertise so the robot learns to reason about terrain while keeping its core gait stable.