aiqbench analysis
This paper presents a humanoid locomotion system that drops the usual motion-capture retargeting, generative motion prior, and distillation stages. It uses a learned five-expert mixture routed by a terrain classifier, paired with a model-based footstep planner refined by gradient steps. The reported gains come from the mixture and refinement components, and the authors' matched ablations show that some intuitive alternatives, such as a single expert or label-free routers, do not explain the results. For AI readers, the work is a useful data point on whether privileged terrain labels and simulation-only evaluation can replace heavier training pipelines, though the results remain limited to one simulator and one robot.
Terrain Traversal Without Motion Priors: Expert-Routed Humanoid Locomotion at Deployment Fidelity
arXiv:2610.14061v1 Announce Type: new Abstract: State-of-the-art perceptive humanoid locomotion is mostly built on retargeted motion capture, a generative motion prior, and a distillation stage. We report a complete system that does not require this pipeline for a humanoid to cross varied terrain. In its place, we provide a five-expert mixture routed by a terrain classifier over a sensed elevation grid, and target every foothold with a model-based planner, refined by six closed-form-gradient steps on an interpolated terrain-cost field instead of picked from a fixed grid. Driven through the same perception-to-actuation software a robot would run, inside a second, deployment-fidelity simulator, it crosses an $85$\,m five-segment waypoint course $18$ of $20$ times end-to-end ($90\%$, Wilson $95\%$ CI $70$--$97$) and succeeds $20/20$ on four of five segments scored alone, where an ablated configuration without the mixture and without refinement reaches $0\%$ on both the stair and the hurdle segment. The continuous refinement moves $72.4\%$ of foot placements across at least one grid-cell boundary, which produces real sub-grid correction. We also ran matched component studies to report what does not account for it: one expert equals five on the training distribution, four label-free routers collapse to near-uniform activation, and a self-supervised terrain embedding loses to the privileged label.
Abstract:State-of-the-art perceptive humanoid locomotion is mostly built on retargeted motion capture, a generative motion prior, and a distillation stage. We report a complete system that does not require this pipeline for a humanoid to cross varied terrain. In its place, we provide a five-expert mixture routed by a terrain classifier over a sensed elevation grid, and target every foothold with a model-based planner, refined by six closed-form-gradient steps on an interpolated terrain-cost field instead of picked from a fixed grid. Driven through the same perception-to-actuation software a robot would run, inside a second, deployment-fidelity simulator, it crosses an $85$\,m five-segment waypoint course $18$ of $20$ times end-to-end ($90\%$, Wilson $95\%$ CI $70$--$97$) and succeeds $20/20$ on four of five segments scored alone, where an ablated configuration without the mixture and without refinement reaches $0\%$ on both the stair and the hurdle segment. The continuous refinement moves $72.4\%$ of foot placements across at least one grid-cell boundary, which produces real sub-grid correction. We also ran matched component studies to report what does not account for it: one expert equals five on the training distribution, four label-free routers collapse to near-uniform activation, and a self-supervised terrain embedding loses to the privileged label.
Robotics (cs.RO)
arXiv:2610.14061 [cs.RO]
(or arXiv:2610.14061v1 [cs.RO] for this version)
https://doi.org/10.48550/arXiv.2610.14061 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 293 words
Related coverage
T-GMP: Terrain-conditioned Generative Motion Priors for Versatile and Natural Humanoid Locomotion
arXiv · Jun 8, 2026
T-GMP: Terrain-conditioned Generative Motion Priors for Versatile and Natural Humanoid Locomotion
arXiv · Sep 2, 2026
GRAIL: Generating Humanoid Loco-Manipulation from 3D Assets and Video Priors
arXiv · Jun 4, 2026
Light-Loco-Parkour: Versatile Perceptive Whole-Body Locomotion via Multi-Skill Distillation
arXiv · Aug 5, 2026