Terrain Traversal Without Motion Priors

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.

Method

System architecture

Everything below the dashed line runs unchanged in both simulators and on a robot. Above it is training-only: a privileged terrain label supervises the gate and, in one configuration, conditions the critic; a self-supervised depth embedding is the alternative critic conditioner compared against it. The deployed actor consumes neither.

No motion-prior dependency

Mixture-of-experts gate

A terrain classifier routes a five-expert mixture over a sensed local elevation grid. The only privileged signal anywhere in the deployed path is the terrain class used to train the gate — no motion-capture data, no generative motion prior.

Model-based, not learned

Continuous foothold refinement

Every foothold target is refined by six closed-form gradient-descent steps on an interpolated terrain-cost field, instead of picked from a fixed grid — real sub-grid correction on 72.4% of placements.

Self-supervised, compared

Terrain-affordance encoder

A temporal-contrastive embedding learned from the robot's own depth stream, with no terrain label. It's evaluated as an alternative critic conditioner — it does not drive the headline transfer result.

Results

Sim-to-sim transfer & component analysis

Chained waypoint course, deployment-fidelity simulator

Per-segment and chained-course success rates for the full system against an ablated configuration.

n=20 per bar, Wilson 95% intervals. The ablated arm removes the expert mixture and the continuous foothold refinement.

Gate routing: privileged label vs. label-free

Gate mass heatmap, privileged-label configuration, near-permutation.

Privileged-label routing (left) is near-permutation; every label-free router tried collapses toward uniform activation.

Stairs (down)

Stairs (up)

Gap crossing

Hurdles

Hurdles (variant 2)

Columns

Rollouts

Deployment-fidelity rollouts, all course segments

Front and back camera for each terrain segment of the chained waypoint course.

No clip yet — drop videos/slope_front.mp4

No clip yet — drop videos/slope_back.mp4

Slope

No clip yet — drop videos/height_field_front.mp4

No clip yet — drop videos/height_field_back.mp4

Height-field

No clip yet — drop videos/random_spread_front.mp4

No clip yet — drop videos/random_spread_back.mp4

Random-spread obstacles

No clip yet — drop videos/up_down_stair_front.mp4

No clip yet — drop videos/up_down_stair_back.mp4

Up / down stairs

No clip yet — drop videos/hurdle_front.mp4

No clip yet — drop videos/hurdle_back.mp4

Hurdles

No clip yet — drop videos/chained_course_front.mp4

No clip yet — drop videos/chained_course_back.mp4

Chained course (all five, in order)

Cite

BibTeX

@inproceedings{anonymous2027terrain,
            title = {Terrain Traversal Without Motion Priors: Expert-Routed
            Humanoid Locomotion at Deployment Fidelity},
            author = {Anonymous Authors},
            booktitle = {IEEE International Conference on Robotics and
            Automation (ICRA)},
            year = {2027},
            note = {Under review}
            }