Collision-Free Is Not Enough: Passage-Conditioned Navigation for Wheeled-Legged Robots
Organizations: Columbia University · Tsinghua University
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.
Figures & tables
| Stored field | Runtime decision it changes |
|---|---|
| Centerline | Defines the candidate coordinates ; every detour is expressed relative to the stored passage location |
| Corridor, detour permission | Bound the candidate offsets and gate the A* fallback |
| Speed cap, approach window | Scale the command and start the slow, short-lookahead approach before a ledge |
| Gait interval | Fixes where a mode request is issued and held |
| Crossing position and direction | Define the frame of the entry gate; a failed check triggers a pivot or a back-up and a new approach |
| Arm (offset injected 0.76 m before the ledge) | Inside both limits | Mean | Mean |
|---|---|---|---|
| Conventional tracking and gait schedule | 33/147 = 22.4% | 13.43 cm | |
| + approach-speed cap | 44/147 = 29.9% | 8.82 cm | |
| + short-lookahead tracking (full approach window) | 70/147 = 47.6% | 5.81 cm | |
| + entry gate | 147/147 = 100% | 0.57 cm | |
| + route-prior window after the ledge | 147/147 = 100% | 0.56 cm | |
| Entry gate alone | 138/147 = 93.9% | 1.13 cm |
| Offset injected | Conventional tracking | + approach window | + entry gate |
|---|---|---|---|
| 2.4 m before the ledge | 95.9% | 100% | 100% |
| 1.6 m | 81.0% | 100% | 100% |
| 1.2 m | 46.3% | 100% | 98.6% |
| 0.76 m | 22.4% | 47.6% | 100% |
| 0.76 m, localization noise (5 seeds) | 24.1% | 46.0% | 88.6% |
| 0.76 m, limits halved to 5 cm and | 9.5% | 31.3% | 100% |
Appendix figures & tables36 assets
Supplementary material from the paper’s appendix.
Appendix
| Speed cap | Tracking | Gate | Prior | Inside limits | Mean (cm) | Mean ( ∘ ) | Holds |
|---|---|---|---|---|---|---|---|
| – | – | – | – | 22.4% | 13.43 | 22.30 | 0 |
| – | – | – | on | 23.1% | 13.42 | 22.21 | 0 |
| – | – | on | – | 93.9% | 1.13 | 3.52 | 2 |
| – | – | on | on | 93.2% | 1.06 | 3.51 | 3 |
| – | on | – | – | 28.6% | 9.76 | 21.64 | 0 |
| – | on | – | on | 27.9% | 9.65 | 21.58 | 0 |
| Back-ups allowed | m | m | m |
|---|---|---|---|
| 0 | 34.7% (30 holds) | 38.1% (27 holds) | 49.7% (16 holds; 29 outside) |
| 1 | 53.7% (23 holds) | 87.8% (7 holds) | 100% (0 holds) |
| 2 | 91.8% (5 holds) | 100% (0 holds), configured | 100% (0 holds) |
| Item | Source | Quantity for the 270.9 m field route |
|---|---|---|
| Centerline, waypoint entries | automatic, from pose trails and marks | 3 full demonstrations (6.1, 6.1, 5.9 min); 29 waypoint entries; 28 segments |
| Local corrections spliced into the line | operator selects recordings; splicing is automatic | 11 re-drives recorded, 5 selected (26, 32, 111, 57, 16 m; 1.5, 0.8, 4.6, 1.5, 0.7 min) |
| Ledge crossing position and direction | automatic, from platform-gait trail events with optional map adjustment | 3 ledges |
| Gait intervals | automatic, from recorded gait events and map steepness | 9 intervals (6 stairs, 3 platform) |
| Compiler settings | operator, one set for the environment | 31 scalar values |
| Runtime settings | shared configuration | 116 values |
| Removed component | Completion | Falls | Obstacle contact | Non-wheel contact |
|---|---|---|---|---|
| local detour | -11.1 | +0.0 | +7.4 | +7.4 |
| entry braking | +7.4 | -11.1 | +0.0 | -3.7 |
| readjustment | -3.7 | +3.7 | -14.8 | -11.1 |
| terrain override | +18.5 | -11.1 | +0.0 | -14.8 |
| Prior reference | Change from prior reference (pp) | |||
|---|---|---|---|---|
| Batch | Completion (%) | Live sensing | Shortest path | Body-aware |
| macOS | 70.4 | -7.4 | -14.8 | -18.5 |
| Linux recheck | 51.9 | +18.5 | +11.1 | +14.8 |
| Linux new seeds | 48.1 | +11.1 | +14.8 | +18.5 |
| Observation | Mean completion effect of a 0.6 m entry lead (pp) |
|---|---|
| Live sensing | +25.0 |
| Route prior |
| Information | Origin | Runtime use |
|---|---|---|
| Centerline and goals | Selected pose trails, passing marks, map-checked geometric processing | Route projection and candidate coordinates |
| Corridor and detour permission | Configured widths and permissions, with clearance checks | Bounds on local alternatives |
| Ledge event and approach pose | Platform-gait trail events, local displacement, optional map adjustment | Event-specific approach tracking and checks |
| Gait intervals | Recorded gait votes, map steepness, configured platform extents | Requested locomotion mode and boundary timing |
| Speed caps and approach extents | Configured route fields and runtime settings | Command limits and activation windows |
| Retry and handoff rules | Shared execution configuration | Re-approach, hold, and switch coordination |
| Time | Ledge stored as | Speed (m/s) | At the first ledge | Run outcome |
|---|---|---|---|---|
| 10:37 | waypoint | 1.0 | crossed at 0.3 m/s in the stairs gait | finished from waypoint 10 (215.6 m) in 572 s, 0.38 m/s |
| 16:53 | waypoint | 1.67 | stalled for 11 s, turn command to rad/s, 0.09 to 0.30 m off the line | paused by the operator |
| 18:33 | waypoint | 1.67 | stalled, turn command at its limit, ended 0.35 m off the line | held |
| 23:18 | event | 1.67 | aligned within 0.05 m; did not mount at 0.4 m/s | held after 20 s without progress |
| 23:36 | event | 1.67 | crossed | log file ends at 160 m |
| 23:52 | event | 1.67 | refused at 0.16 m, one back-up, crossed within 0.07 m | 29 of 29 waypoints in 386.6 s, no hold |
| Rejoining | Entry conditioning | Control | Intended displacement |
|---|---|---|---|
| Off | Off | 20.8 | 0.0 |
| Off | On | 29.2 | 0.0 |
| On | Off | 25.0 | 37.5 |
| On | On | 45.8 | 50.0 |
| Previous (3 seeds) | All 23 seeds | |
| Qualifying route–seed sets / matched pairs | 16 / 32 | 113 / 226 |
| Ordered completion, rejoining off | 0% | 0% |
| Ordered completion, rejoining on | 59.4% | 57.1% |
| difference, 95% seed-block bootstrap | 50.9 to 63.4 points | |
| Ordered completion, same sets undisplaced, rejoining on | 43.8% | 54.0% |
| Mean terminal distance, rejoining off / on | 1.14 m / 0.11 m | 1.10 m / 0.14 m |
| Segment | Cells | Hold | Turn and side-step | Walk and steer |
|---|---|---|---|---|
| WP01–04 | 46 | 0 | 33 | 32 |
| WP05–08 | 19 | 0 | 0 | 10 |
| WP09–12 | 38 | 0 | 18 | 18 |
| WP17–20 | 46 | 0 | 3 | 18 |
| WP21–23 | 46 | 0 | 28 | 38 |
| WP24–26 | 46 | 0 | 25 | 21 |
| Displacement | Sets | Hold | Rejoin | Undisplaced | Rejoin undisplaced (95%) | Returned / left |
|---|---|---|---|---|---|---|
| 0.3 m | 117 | 34.6 | 49.6 | 53.0 | ( to ) | 143 / 145 |
| 0.6 m | 117 | 32.5 | 53.4 | 53.0 | ( to ) | 136 / 136 |
| 0.9 m | 115 | 1.7 | 52.6 | 53.0 | ( to ) | 204 / 204 |
| 1.2 m | 113 | 0.0 | 57.1 | 54.0 | ( to ) | 226 / 226 |
| 1.5 m | 109 | 0.0 | 29.8 | 55.0 | ( to ) | 114 / 218 |
| 1.8 m | 78 | 0.0 | 0.0 | 56.4 | ( to ) | 0 / 156 |
| Completion (%) | Touching the box | Ending in a hold | |||||
|---|---|---|---|---|---|---|---|
| Box offset | off | on | On no box (95%) | off | on | off | on |
| no box | 33.7 | 57.6 | 0 | 0 | 65 | 2 | |
| 0 m | 6.0 | 9.2 | ( to ) | 6 | 154 | 161 | 36 |
| 0.3 m | 35.9 | 65.2 | ( to ) | 0 | 6 | 79 | 0 |
| 0.6 m | 34.8 | 60.9 | ( to ) | 0 | 0 | 79 | 2 |
| 0.9 m | 39.7 | 57.6 | ( to ) | 0 | 0 | 65 | 0 |
| Ordered completion | Centre within 0.20 m | Body touches the 0.25 m disk |
|---|---|---|
| 1.2 m displacement, hold (226 runs) | 0.0% | 0.4% |
| 1.2 m displacement, rejoining | 57.1% | 75.7% |
| Same sets, not displaced | 54.0% | 77.4% |
| Three behaviours, 241 cells: hold | 0.0% | 0.4% |
| default turn and side-step | 44.4% | 61.4% |
| walking forward while steering | 56.8% | 75.9% |
| Heading error at handoff | Passed | Rate | Lateral error at handoff | Passed | Rate |
|---|---|---|---|---|---|
| to | 74 / 75 | 98.7% | 0 to 0.05 m | 52 / 52 | 100% |
| to | 65 / 70 | 92.9% | 0.05 to 0.10 m | 71 / 71 | 100% |
| to | 51 / 66 | 77.3% | 0.10 to 0.20 m | 52 / 53 | 98.1% |
| to | 42 / 63 | 66.7% | 0.20 to 0.30 m | 41 / 46 | 89.1% |
| to | 29 / 46 | 63.0% | 0.30 to 0.60 m | 13 / 13 | 100% |
| to | 16 / 49 | 32.7% |
| Stairs zones | Zones | Handoffs | Passed | Falls | Recover or hold |
|---|---|---|---|---|---|
| First riser crossed at to | 4 | 181 | 164 (90.6%) | 8 | 9 |
| First riser crossed below , or no riser | 11 | 474 | 468 (98.7%) | 2 | 4 |
| All | 15 | 655 | 632 (96.5%) | 10 | 13 |
| Displaced 0.76 m before each zone | Alignment on | Alignment off |
|---|---|---|
| Climbs passed | 180 / 180 (100%) | 171 / 180 (95.0%) |
| Falls | 0 | 9 |
| Median heading error at the handoff | ||
| Handoffs within | 180 / 180 | 37 / 180 |
| Zone reached at about without alignment | 36 / 36 | 27 / 36 |
| Other four zones | 144 / 144 | 144 / 144 |
| Heading error at entry | Entries | Climbed | Fell | Stalled |
|---|---|---|---|---|
| to | 675 | 57.9% | 8.9% | 33.2% |
| to | 900 | 60.8% | 9.9% | 29.3% |
| to | 900 | 55.0% | 12.6% | 32.4% |
| 450 | 50.4% | 18.4% | 31.1% | |
| Inside the gate’s bound (10 cm, ) | 225 | 61.8% | 5.3% | 32.9% |
| Outside the bound | 2,700 | 56.3% | 12.3% | 31.3% |
| Reference + prior | Live sensing | Shortest path | RB-TRG core | |||||
|---|---|---|---|---|---|---|---|---|
| Task | C | F | C | F | C | F | C | F |
| Flat | 87 | 0 | 87 | 0 | 88 | 0 | 72 | 0 |
| Stairs | 18 | 72 | 67 | 16 | 76 | 10 | 41 | 56 |
| Rock approach | 58 | 36 | 30 | 63 | 8 | 91 | 8 | 89 |
| Platform | 45 | 42 | 58 | 23 | 42 | 46 | 48 | 44 |
| Terrace | 5 | 91 | 58 | 34 | 12 | 84 | 17 | 80 |
| Scale | Task | Prior | Live | Shortest | RB-TRG |
|---|---|---|---|---|---|
| 0.8 | wp05_08_stairs | 2 | 5 | 7 | 1 |
| 0.8 | wp17_21_platform | 7 | 2 | 4 | 2 |
| 0.8 | wp26_29_terrace | 1 | 1 | 5 | 6 |
| 0.8 | wp28_30_descent | 3 | 7 | 6 | 9 |
| 1.0 | wp05_08_stairs | 3 | 5 | 5 | 1 |
| 1.0 | wp17_21_platform | 5 | 2 | 4 | 5 |
Robot Navigation
Robot navigation studies how robots perceive, localize, plan, and move safely toward goals in static or changing environments.
See all 595 papers