Terrain-Indexed Motion Priors and Geometric Support for Extreme Humanoid Parkour

Chenghao Sun, Youhao Lyu, Zizheng Li, Shuting Zhang, Ziyang Dong, Wenhao Cui

November 2026

TL;DR One depth-based humanoid policy learns to climb and cross obstacles up to about 1 m tall from human motion clips, then is retrained to place its hands and knees where they suit the robot. Deployed zero-shot on a Noetix E1, it clears 29 of 30 single obstacles and 4 of 5 three-obstacle courses.

What they do

Human parkour clips show how to get over an obstacle, but a robot with different proportions may need its hands and knees elsewhere. One perceptive humanoid policy is trained in three stages: make human clips executable, learn terrain-appropriate skills under motion priors, then relax the priors and reward support that suits the robot. It picks skills from onboard depth and proprioception, with no skill label.

How it works

Reference preparation. Privileged mimic training turns eight retargeted human clips into executable robot rollouts with torques, contacts and height scans.
Terrain-indexed motion priors. An AMP discriminator with a shared trunk and four heads (flat, and Low, Mid and High frames of about 0.4, 0.6 and 1.0 m) scores the policy against the current terrain’s references. The actor sees a 24x24 depth image, proprioception and velocity commands.
Support adaptation. Flat, reachable patches on obstacle surfaces become hand and knee candidates; rewards pull limbs toward them and score the support achieved, while the AMP weight is halved.

Results

Simulation (Isaac Lab, 2,048 episodes per condition): 96.3% success on a Low-Mid-High course vs 79.7% for a multi-skill distillation baseline; a single AMP prior and plain velocity tracking never clear High. At heights 1-3 cm outside training: 93.5% vs 65.3%.
On a 1 m tall, 1 m x 1 m platform: 93.5% success, vs 46.2% without candidate rewards and 58.9% without relaxing the prior.
Zero-shot on the 23-DoF Noetix E1: 10/10 Low, 9/10 Mid, 10/10 High, 4/5 courses, 13/15 at out-of-range heights.

Limitations

On hardware, the failed Mid trial lost arm support, and the failed course had an abrupt heading change followed by the Mid skill on a Low platform; the authors name support maintenance and skill transitions as open problems. Stage ablations also drop training iterations, so they do not isolate each stage’s objective. Candidates are near-flat patches; irregular surfaces are future work.

Figures

From Figure 2: Training framework. Retargeted human clips are refined into physically executable robot references (ri... — From Figure 2: Training framework. Retargeted human clips are refined into physically executable robot references (right); terrain-indexed discriminators compare reference and policy motion windows for a style reward (top); hand and knee support candidates on obstacle surfaces drive placement and contact rewards (bottom). Only the actor and its observation encoders are deployed.

From Figure 5: Hand and knee support before (top) and after (bottom) the support-adaptation stage, in simulation (lef... — From Figure 5: Hand and knee support before (top) and after (bottom) the support-adaptation stage, in simulation (left) and on hardware (right). Before adaptation, hand contacts sit near the platform edge (red circles).

From Figure 4: One policy crossing three obstacles in a row in simulation (top). The joint-angle DTW distance to the ... — From Figure 4: One policy crossing three obstacles in a row in simulation (top). The joint-angle DTW distance to the High, Mid and Low reference clips (middle) shows the closest reference switching from High to Mid to Low (bottom).

Abstract

We present a progressive imitation framework that learns terrain-appropriate humanoid parkour skills and subsequently adapts their hand--knee support. Human demonstrations provide traversal behaviors, but their contact patterns need not suit a robot's proportions or the available support surfaces. Our framework changes the role of demonstrations across three stages: privileged imitation produces physically executable robot references; terrain-indexed adversarial motion priors guide skill acquisition; and geometric support guidance with a reduced motion-prior weight enables contact adaptation. Candidate-based precontact attraction and actual-contact evaluation guide support without prescribing a contact sequence. The deployed policy selects behaviors from onboard depth, proprioception, and velocity commands without an external skill label. On the Noetix E1, the policy achieves 96.34\% hybrid success in simulation and transfers without real-world policy fine-tuning, completing 29 of 30 individual obstacle trials and four of five hybrid courses. Support-adaptation ablations show higher traversal success for the complete adaptation stage on the tested terrains, while qualitative observations show changes in contact placement. These results support acquiring skills from demonstrations before relaxing their influence to accommodate robot-specific support configurations.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong