← Paper Digest

Collision-Free Is Not Enough: Passage-Conditioned Navigation for Wheeled-Legged Robots

Wai Ka Sun, Waiyiu Pong, Bowen Wang

November 2026

TL;DR Store each taught stair and ledge crossing as an event on the route, and make a wheeled-legged robot back up and retry before a ledge unless it is within 10 cm and 10° of the taught entry. Under localization noise this raises in-limit ledge crossings from 24.1% to 88.6%; on the real robot it completed all six runs of a 271 m competition course.

What they do

Wheeled-legged robots cross stairs and ledges with fixed locomotion skills that succeed only from a limited set of entry states. A collision-free detour can deliver the robot to a ledge a few centimetres or degrees off, and the skill stalls or falls. Passage-conditioned navigation takes the entry condition from one teaching session (27 minutes of driving) instead of a robot model or extra trials, and enforces it after every local replan. Nothing is trained on site.

How it works

  • Stored route. Teaching pose trails, waypoint marks and gait events compile into a route of 270.9 m with 29 waypoints. Each stairs or ledge transition is an event placed by distance along the route, with a gait interval and a speed cap; each ledge also keeps the position and direction of its taught crossing.
  • Approach window and entry gate. Near a ledge the robot slows and tracks the stored line closely. 0.7 m before the crossing it checks its pose in the crossing frame; if it is more than 0.10 m or 10° off, it turns in place or backs up and approaches again.
  • Local repair. A LiDAR local planner detours around unmapped obstacles inside a corridor; a displaced or stalled robot walks forward while steering back to the stored line.

Results

  • Kinematic simulation, offsets injected 0.76 m before each ledge: conventional tracking keeps 22.4% of crossings within both limits, the approach window 47.6%, the gate 100%. Under localization noise: 24.1%, 46.0% and 88.6%. Time from offset to crossing rises from 4.2 s to 11.3 s.
  • Contact simulation with learned policies: no falls in 180 stair climbs after a push, against 9 when the stored direction is ignored. After a 1.2 m push, 75.7% of segments are completed with rejoining (0.4% without), matching undisturbed runs (77.4%).
  • Real robot with factory gaits: all six competition-day runs of the route, which spans 7.0 m of elevation, completed autonomously, placing third in its task.

Limitations

The gate is conservative: of approaches made without it that fail its check, 50.8% still crossed within the limits. Skill choice and stepping over obstacles are explicit hand-set rules. On stairs and around ledges the planner uses the stored route instead of live sensing, which covers 58% of the route. The simulation studies use the authors’ learned policies, not the robot’s factory gaits.

Figures

From Figure 1: (a) One teaching session and the situations after it. (b) The stored route conditions motion on the li...
From Figure 1: (a) One teaching session and the situations after it. (b) The stored route conditions motion on the live scan, and an entry gate checks every ledge approach. (c) Ledge crossings inside both entry limits as stored fields are added. (d) Outcomes after a 1.2 m push. (e) The competition route.
From Figure 3: (a) Teaching, the map and configured rules compile once into the stored route, which a body sweep and ...
From Figure 3: (a) Teaching, the map and configured rules compile once into the stored route, which a body sweep and a full kinematic run check before use. (b) At every control step the stored route, the local observations and the candidate motions share one route frame; the robot compares the stored entry with its live pose. (c) The decisions of one step, with excerpts of the robot log.
From Figure 6: Crossing errors at the three ledges with ledge handling disabled and enabled. Top: for each injected l...
From Figure 6: Crossing errors at the three ledges with ledge handling disabled and enabled. Top: for each injected lateral and heading offset, the worst normalized error among the three ledges; 1 or less means both the 10 cm and 10 degree limits are met. Bottom: cumulative distributions of the absolute lateral and heading errors.
From Figure 10: Video frames of the robot at a ledge: the approach, the repositioning, the second approach and ascent...
From Figure 10: Video frames of the robot at a ledge: the approach, the repositioning, the second approach and ascent, and the traversal that follows.
Abstract

A collision-free detour can still bring a legged robot to the next stair or ledge a few centimetres or degrees off, and a fixed locomotion skill, factory or learned, can stall or fall. Existing methods prevent this with a robot model or extensive trial and error, but on a new site with unknown skills neither is available, so they cannot be used. We find that a single 27-minute teaching session already supplies what these methods have to derive or learn: for every stair and ledge, one crossing the skill has completed. Our method, passage-conditioned navigation, makes the route remember how each transition was entered, not only where the path went. It stores each taught crossing as an event and, without training or a robot model, enforces it after every local replan: before each ledge, an entry gate makes the robot turn or back up and retry if it is more than 10 cm or 10° off the taught crossing. Under localization noise, with path, gait schedule, and sensing unchanged, this raises the share of disturbed ledge crossings that stay within these limits from 24.1% to 88.6%. In contact simulation with our learned policies, a robot pushed off course before stairs never falls in 180 climbs, against 9 falls when the stored direction is ignored; pushed 1.2 m away, it rejoins the route, ending on average 0.14 m from it (1.10 m without rejoining), and finishes as often as an undisturbed robot. This suits any robot with fixed skills that repeats a route over terrain transitions, as in inspection and patrol, where one wrong entry can end the task. We deployed it in competition on a real wheeled-legged robot: over 271 m of multi-obstacle terrain spanning 7.0 m of elevation, it completed all six runs, won third place at the 2026 Global Open-source AI Challenge, which drew 2,999 submissions, and was the only finalist in its task to finish fully autonomously.

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.