acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

A collision-free detour can still leave a robot unable to negotiate the next stair or ledge. For wheeled-legged robots, negotiating terrain depends on approach geometry, speed, and gait as well as obstacle clearance. We introduce passage-conditioned navigation: a persistent reference combines demonstrated geometry and gait events with configured rules, while observations and execution feedback guide local adaptation with a fixed locomotion library. On a teaching-compiled route, event-conditioned handling reduces mean lateral entry error from 13.43 to 0.56 cm across 98 deterministic kinematic rollouts. In matched contact-simulation perturbation cases across six segments of a separate, frozen mapping-derived reference, rejoining reduces mean terminal route deviation from 1.14 m to 0.11 m and raises ordered completion from 0% to 59.4%. Across 2,800 contact-simulation trials, the live-sensing reference achieves 62.6% completion, exceeding two graph-based route alternatives by 18.6 and 19.0 percentage points with the same executor. An outdoor deployment traverses successive terrain transitions using factory gait controllers; the navigation stack also runs on an Apple M1 laptop. A geometry stress test exposes limits to transfer. Together, these results show how stored passage conditions and reference-based recovery support local adaptation, and where controller and observation interactions still constrain reliability.

Primary area:
applications to robotics, autonomy, planning
Keywords:
wheeled-legged robots, robot navigation, teach-and-repeat, local replanning, route recovery, terrain transitions

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_TGv8OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

In the paper "Collision-Free Is Not Enough: Passage-Conditioned Navigation for Wheeled-Legged Robots," the authors propose a navigation framework that preserves information from previously successful trajectories, including approach geometry, speed limits, and locomotion modes, while allowing local replanning around obstacles. They demonstrate the approach on a wheeled-legged robot and report improved task completion and trajectory recovery in simulation.

Overall, I find that the paper lacks a clear methodological contribution. The underlying problem is reasonable, but it is unclear how the proposed approach differs fundamentally from existing reference-based or kinodynamic planning methods. Much of the proposed framework appears to combine standard trajectory optimization, reference tracking, and manually specified execution constraints. The paper does not convincingly establish what new capability this combination provides.

Critical strengths

The paper is supported by a physical robot deployment across multiple terrain transitions, along with extensive simulation experiments. The authors also provide component ablations and report negative results, which help characterize the limitations of their implementation.

Critical weaknesses

My primary concern is the lack of a clearly defined methodological contribution. The authors argue that collision avoidance alone is insufficient because successful traversal also depends on approach geometry, velocity, and locomotion mode. However, these considerations are already central to kinodynamic motion planning. It is unclear what specific limitation of existing methods the proposed framework addresses.

The proposed method appears to retain a demonstrated reference trajectory, associate it with manually configured constraints, and perform local replanning while encouraging the robot to return to that reference. These are established planning and control mechanisms. The paper does not clearly explain what is novel about their integration or why existing formulations could not accomplish the same objective.

The experimental results also provide limited evidence for the proposed contribution. The strongest improvements come from reference rejoining and explicit approach correction, both of which are expected behaviors of conventional trajectory-following methods. The experiments do not convincingly isolate the benefit of the proposed passage representation.

Questions for authors
  1. What distinguishes the proposed approach from existing kinodynamic planning or reference-based replanning with maneuver-specific constraints?

  2. How are the passage conditions determined, and what evidence establishes that preserving them improves maneuver feasibility?

  3. Can the authors demonstrate an improvement attributable specifically to passage conditioning, independent of conventional reference rejoining, approach correction, and gait scheduling?

6 comments
AuthorsOpenReview ↗

Response to Reviewer TGv8 (1/2): contribution and Q1

We sincerely thank the reviewer for the careful assessment and for three clear questions, which helped us sharpen the paper. In short, the revision responds to each point as follows.

  • Methodological contribution: now stated as three contributions in the revised introduction (below).
  • Q1, the difference from kinodynamic and reference-based planning: answered below, with extended related work (Section 2).
  • Q2, how the conditions are determined and the evidence: second comment (Table 10, Section 5.2, Appendix N).
  • Q3, an improvement attributable to passage conditioning alone: a new factorial experiment (Section 5.2; second comment).

Numbers refer to the revised PDF.

The contribution. We may not have made it clear enough in the previous version, and the revised introduction now states it directly. The paper addresses one specific limitation: when the locomotion controllers are fixed and can be neither modelled nor retrained, a planner has no description of the states from which a stair or ledge can be climbed. We are indebted to earlier work on route following, kinodynamic planning and skill composition; building on what it explored, the paper contributes three things:

  1. it identifies what has to be stored for each transition (a crossing frame, an approach window and a gait interval), all obtained from one teaching session;
  2. it gives an interface that enforces this over a black-box skill library after any local replan;
  3. it measures what each stored field adds over conventional tracking of the same path: 22.4% of disturbed ledge crossings inside the entry limits with conventional tracking, 47.6% with the stored approach window, and 100% with the entry gate, in kinematic simulation (Section 5.2).

Q1. Difference from kinodynamic planning and reference-based replanning with maneuver-specific constraints. The difference lies in what is assumed to be available. We agree that approach geometry, speed and locomotion mode are classical concerns of kinodynamic planning, and the revised related work says so. A kinodynamic or maneuver-constrained planner derives the admissible approach states from a model of the system. Our controllers are black boxes (factory gaits on the robot; in simulation, two PPO policies that we trained and then froze), so there is no such model.

The learning literature obtains these states through further interaction or through access to the skills: skill chaining learns initiation sets from repeated interaction (Konidaris & Barto, 2009; Bagaria & Konidaris, 2020), terminal-state regularization shapes where each policy ends during training (Lee et al., 2022), STAP scores a skill sequence with the skills' value functions (Agia et al., 2023), and ANYmal parkour trains a navigation policy over its locomotion skills (Hoeller et al., 2024). Closest to our setting, Hedegaard et al. (2025) plan with black-box skills by inserting motion plans that bring the robot to a configuration from which each skill can start. We study the case in which those start conditions are not given: they are recorded in one teaching session in which every transition was crossed.

Reference-based replanning (e.g., Sehn et al., 2022) retains the taught path through a deviation penalty. The new experiment tests directly whether retaining the path is enough, and within about 1.2 m of a transition it is not (Table 3). We retain, in addition, one event per transition and enforce it where it occurs. These references are now in the paper.

Practical evidence. The system was developed for a public competition in which every team could draw on the same published wheeled-legged, driving and aerial navigation methods, and we started from that literature ourselves; we found no open-source method that met the conditions of the task together. Our first versions followed a fixed route and failed under small changes in the scene and under map and localization error; in contact simulation, route following completed the full route in 4 of 24 seeds without recovery and in 19 of 32 with it (Appendix N). The final system won third place at the competition, which drew 2,999 submissions, and was the only finalist in its task to complete the route fully autonomously (Section 5.5). We also looked at commercial navigation services. We have not used them, but their public documentation describes related functions, for example route edges annotated for stairs in Boston Dynamics' GraphNav; they are proprietary and tied to the vendor's robots, and we found no published account of how they enforce entry. Our method needs only one simple teaching session, and the navigation stack and benchmark are released with the paper.

We would be glad to discuss any specific formulation the reviewer has in mind.

AuthorsOpenReview ↗

Response to Reviewer TGv8 (2/2): Q2 and Q3

Q2. How the passage conditions are determined, and the evidence that preserving them improves feasibility. They come from the demonstration: the crossing position, the crossing direction and the gait intervals are taken from the teaching recordings. The tolerances (10 cm, 10°), window lengths, speed caps and retry count are settings shared by the whole environment (Section 4.1, Table 10). The evidence is of three kinds.

  • Robot (Table 11): with the ledge stored as an ordinary waypoint, both full-speed runs stalled at the first ledge. With the ledge stored as an event, the gate refused an approach 0.16 m off the line, the robot backed up once and crossed within 0.07 m, and the run reached all 29 waypoints. These are consecutive versions of one system, and the controlled comparison is the factorial below.
  • Learned skill in contact simulation (Appendix N): our stairs policy passes 98.7% of entries within 5° of heading error and 32.7% beyond 25°.
  • Controlled experiment (Section 5.2): the factorial under Q3.

Q3. An improvement attributable to passage conditioning, independent of rejoining, approach correction and gait scheduling. The new experiment was designed around this question (Section 5.2). Every arm has the same centerline, gait schedule, sensing input and feedback model, and the first arm is conventional tracking with gait scheduling, so approach correction by a conventional tracker is the baseline. Share of disturbed ledge crossings inside the entry limits:

Disturbance arrivesConventional tracking + gait schedule+ stored approach window+ entry gate
2.4 m before the ledge95.9%100%100%
1.2 m46.3%100%98.6%
0.76 m22.4%47.6%100%

A conventional tracker corrects the approach when it has distance (95.9% at 2.4 m) and does not in the last 1.2 m, which is where a detour that rejoins late leaves the robot. The part attributable to the stored events is that remainder: +25.2 points for the approach window at 0.76 m, and the rest for the gate, whose main effect over all other factor combinations is +64.5 points. The gate column is measured on entry geometry in kinematic simulation: the gate refuses an approach outside the limits and the robot approaches again, and no rollout at 0.76 m ends in a hold. Its check is also conservative: across 1,470 approaches made without the gate, all 350 that pass the check end inside the limits, against 50.8% of those that fail it.

On rejoining, we separated the effect of returning to the reference from the way the robot returns, by adding the tracker's default recovery behaviour (turn and side-step) as a plain baseline to our strategy (walk forward while steering to the stored line; Section 5.3). Runs are scored by the competition's waypoint rule: the body touches the 0.25 m disk of every goal in order, and a fall voids the run. After a 1.2 m push, completion is 0.4% when the robot holds, 61.4% with the default behaviour and 75.9% with our rejoining strategy (14.5 points or 24% more; interval 8.2 to 21.3), against 78.0% for the same cells with no push. Under the stricter rule of the previous version (centre within 0.20 m, a target 36% smaller in area) the three figures are 0%, 44.4% and 56.8%. Our strategy returns the robot to the undisturbed level and reaches the next goal in all 241 cells, a median of 1.6 s later than without a push; the default behaviour reaches it in 216 cells, a median of 5.8 s later.

Two points that the previous version may not have made clear enough. The method contains no trajectory optimization: Section 4.2 generates a small set of lateral offsets in route-relative coordinates, scores them with the closed-form cost of Eq. (5), and falls back to a bounded A* search where detours are permitted. And the constraints are configured by hand only in part: the limits are set once for the whole environment, while the crossing position, the crossing direction and the gait intervals come from the demonstration (Table 10). The revised Section 4 states both.

We sincerely hope that these clarifications and the new experiment address the concern about the contribution, and that the reviewer would consider raising the score. Thank you again for the thoughtful review; we wish you every success in your research and would welcome further exchange.

AuthorsOpenReview ↗

Further response to Reviewer TGv8: Q1 from practice

We thank the reviewer again and add one point to Q1, from practice. Before this method, we tried the standard approaches in the same environment. Planning the route from the map with two grid planners of our own, one over flat cells and one over cells and gaits, found a path for 13 of 29 segments (Appendix B). Following a fixed route without recovery completed the full route in 4 of 24 contact-simulation seeds (Appendix N). On the robot, on the day before the final, earlier versions that followed the route without the re-adjust state ended 9 of the 13 logged runs in a hold: 4 when the robot was more than 0.8 m off the route and 5 after 8 s without progress. With the ledges stored as ordinary waypoints, both full-speed runs stalled at the first ledge (Appendix I, Table 11). In contact simulation, replacing the taught centerline inside the same corridor by shortest-path planning or by the RB-TRG body-aware planner gave 44.0% and 43.6% completion, against 62.6% for the taught reference with live sensing (Section 5.4). The final version, which stores each transition as an event and rejoins the reference, completed that evening's full run without a hold and six of six runs on the competition day.

The robot record of the earlier versions is now stated in Appendix I of the updated PDF. We would be glad to clarify anything further.

Reviewer_47rhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes a passage-conditioned navigation framework for wheeled-legged robots that combines stored route experience with online local adaptation. The paper addresses the limitation that a collision-free detour may compromise the approach geometry, speed, or gait needed for a subsequent terrain transition. The authors compile demonstrated pose trails and gait events, map geometry, and configured rules into a persistent reference containing route corridors, passage conditions, speed limits, and gait intervals. During execution, route-relative planning uses local observations to generate candidate paths, while maintained runtime state coordinates obstacle avoidance, reference rejoining, and confirmed gait transitions. The framework also selects between live terrain observations and stored route priors to traverse familiar terrain with an existing locomotion library. Experiments support improved approach alignment and route recovery in the evaluated settings, while revealing limitations in transfer across terrain and execution conditions.

Critical strengths
  1. The framework addresses a practical limitation of collision-based planning: avoiding an obstacle does not ensure a suitable approach to the next terrain transition. Linking passage geometry and execution conditions to a shared route coordinate provides a concrete interface for local adaptation with existing locomotion controllers. The teaching-to-reference correspondence makes the origins of the stored information explicit.

  2. The controlled comparisons and transparent outcome reporting make the empirical results informative. Shared controllers and execution settings help assess route preparation and observation selection, while paired analyses support the reported differences. Retaining falls, invalid setups, negative component effects, and counterexamples allows readers to evaluate the limits of the approach alongside its successes.

Critical weaknesses
  1. The core methodological contribution is unclear. Compiling reference fields and maintaining runtime state describe system functions, but the distinct planning mechanism is not well established. Section 4.2 mainly scores clearance and route-relative motion, while entry conditions are checked during execution in Section 4.3. The paper does not clearly explain how those conditions constrain local replanning, and the coupled comparisons leave their added value over conventional route following and gait scheduling unresolved.

  2. The preparation cost and generality of the approach are insufficiently established. The system relies on teaching recordings, operator-selected corrections, and configured passage rules, but the effort and benefit relative to map-based planning are not evaluated. Results are largely restricted to one reconstructed course, with important comparisons reversing across settings. The outdoor deployment uses different controllers without a matched comparison, limiting evidence of transferable benefit.

Questions for authors
  1. How do the stored passage conditions change the local planning decision? The candidate cost in Section 4.2 primarily addresses obstacle clearance, route-relative offsets, and switching stability, whereas approach checks and gait coordination appear in the execution logic of Section 4.3. Please identify the specific constraints, costs, or candidate-generation rules through which the next passage's position, heading, speed, and gait requirements affect the selected detour. A comparison using the same centerline and executor with conventional tangent tracking and gait scheduling would help establish what the proposed representation adds. The current graph comparisons retain several passage fields, while the kinematic diagnostic changes multiple handling mechanisms together, making this distinction difficult to assess.

  2. What preparation is required to obtain a usable reference for a new route? The method includes teaching recordings, selected trajectory corrections, and configured corridor widths, approach extents, speed limits, and tolerances. Please distinguish automatically derived fields from operator decisions and report any available information about the required teaching and configuration effort. It would also help to explain which settings transfer across routes or terrain changes and what requires a new demonstration or manual adjustment. Since the main tasks are subroutes of one reconstructed course and performance comparisons vary across execution settings, the practical benefit over planning from a map and goal sequence remains difficult to judge.

  3. I would appreciate more detail on how the system handles changes that invalidate a stored passage prior. When route-profile observations replace live terrain evidence, what current signals can still stop the robot for a newly introduced obstacle? Section 4.3 also describes relaxed retry and late-arrival branches that may allow crossing without the nominal heading check. Please explain when these exceptions are appropriate and how their tolerances were selected or validated. These details are important for assessing whether the stored conditions remain executable after local replanning.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 47rh (1/2): Q1 and Q2

We sincerely thank the reviewer for the detailed questions, which pointed to several places where the previous version was unclear and led to concrete changes. In short:

  • Q1, how the stored conditions change the planning decision: a new Table 1, two candidate-generation rules, and the comparison the reviewer proposed (Section 5.2).
  • Q2, preparation for a new route: itemized below and in Table 6, with the benefit over planning from a map (Section 5.1).
  • Q3, changes that invalidate a stored prior: next comment, with a new experiment on the exception branches (Table 5).

Numbers refer to the revised PDF.

Q1. How the stored passage conditions change the local planning decision. They act in three places: in how candidates are generated, in the approach, and in a check before the crossing. The reviewer's reading of Eq. (5) is correct: the cost scores clearance, route-relative offset and switching stability, and the entry pose is not a term of it. The previous version did not state this plainly, and the revision adds Table 1:

Stored fieldRuntime decision it changes
CenterlineDefines the candidate coordinates (s, d); every detour is expressed relative to the stored passage location
Corridor, detour permissionBound the candidate offsets d_j and gate the A* fallback
Speed cap, approach windowScale the command and start the slow, short-lookahead approach before a ledge
Gait intervalFixes where a mode request is issued and held
Crossing position and directionDefine the frame of the entry gate; a failed check triggers a pivot or a back-up and a new approach

Two candidate-generation rules depend on the next passage. Where the stored route does not permit a detour, for example on stairs, the only candidate is the stored line. And when the next goal lies within the 3.5 m planning horizon, every candidate must bring its lateral offset back to zero over at most the last metre before it, so the next passage bounds where a detour may end. Heading and speed at the transition are enforced afterwards, by the approach window and the entry gate.

The comparison the reviewer proposes, the same centerline and executor under conventional tangent tracking and gait scheduling, is the first arm of the new kinematic factorial in Section 5.2. With a disturbance 0.76 m before a ledge it brings 22.4% of crossings inside the entry limits; the stored approach window brings 47.6% and the gate 100%. The four mechanisms that the previous version's diagnostic changed together now have separate switches.

Q2. Preparation required for a new route. Table 6 (Appendix B) lists the preparation; the version below adds detail on the settings.

ItemSourceQuantity for the 270.9 m field route
Centerline, waypoint entriesautomatic, from pose trails and passing marks3 full demonstrations (6.1, 6.1, 5.9 min); 29 waypoints; 28 segments
Local corrections spliced into the lineoperator selects recordings; splicing is automatic11 re-drives recorded, 5 selected (0.7 to 4.6 min each)
Ledge crossing position and directionautomatic, from platform-gait trail events with optional map adjustment3 ledges
Gait intervalsautomatic, from recorded gait events and map steepness9 intervals (6 stairs, 3 platform)
Compiler settings (corridor widths, clearances, lane and platform extents)operator, one set for the environment31 scalar values
Runtime settings (entry limits, retry rules, speed caps, gains)shared configuration116 values

Effort. One teaching session: the operator drives the route two or three times at about six minutes per pass, and re-drives a passage where it needs a better line. The one-time map recording took 25 minutes. Nothing is trained and no parameter is fitted, and autonomous runs on the compiled reference began the same day.

What transfers. The contact-simulation experiments and the field stack share the navigation library byte for byte (35 source files) and 106 of the 116 runtime values (Appendix B). A new route needs new demonstrations and a recompile; a terrain change at an existing ledge needs a new recording of that passage; corridor widths and tolerances are configuration edits.

Benefit over planning from a map and a goal sequence. We tried this on the same venue before the teaching compiler existed, with two grid planners of our own (Section 5.1, Appendix B). On the 29 segments, the better planner returned a path for 13, and for 4 with a 0.4 m body clearance, mostly because a waypoint cell itself was not passable in the map. No setting gave a usable route across all 29 segments. A taught reference does not depend on map quality at every waypoint, adapts during the run (detours, the entry gate, rejoining), and keeps changes local: a changed waypoint needs one re-drive of that passage and no new map.

AuthorsOpenReview ↗

Response to Reviewer 47rh (2/2): Q3

Q3. Changes that invalidate a stored prior, and the exception branches. The answer depends on where the change occurs, so we give the three cases in turn.

Where live sensing is active (58% of the field route). The local planner detours around obstacles that are absent from the map and then continues the task. During the deployment this included team members standing beside the final approach: the robot sensed them, kept clear and continued to the finish, which is recorded on video (not included in the anonymized supplement). In contact simulation we also inserted an unmapped box (Appendix L). On open stretches, with rejoining, a box that takes part of the corridor is passed: the robot reaches the next goal in 366 of 368 runs and touches the box in 6 (182 of 184 with no box). A box that blocks the corridor, or stands where the stored route permits no detour, is not passed: without rejoining the robot holds, and with rejoining it pushes on and touches the box, which we report as a limit.

Inside the route-prior intervals. Route-prior observations replace the live grid inside the stairs and platform-gait intervals and in the 1.5 m that follow each ledge's platform-gait interval, all of which are traversed inside a narrow corridor. To answer the question directly: inside those intervals the planner does not see a newly introduced obstacle, and the signals that remain live are route progress and distance from the line. A robot that stalls or leaves the line rejoins the reference by walking forward while steering back to the stored line. In the logged end-to-end run this happened four times, outside the ledges; each time the robot resumed within 2 s, and it did not enter the hold state. A hold is the safety fallback, entered after 20 s of rejoining without 0.5 m of progress or beyond 1.5 m from the line (Section 4.3). A stale pose (0.3 s) or observation (0.5 s) also stops the follower (Appendix K).

The exception branches. The relaxed-retry branch keeps the robot from holding indefinitely at a ledge whose passable lane is wider than the nominal 0.10 m tolerance, and the late-arrival branch covers a start beyond the check point. The retry count and the relaxed bound were set on the field platform before the simulation experiments. Their effect is now measured (Table 5; 441 further rollouts with 0, 1 or 2 back-ups and a relaxed bound of 0.15, 0.25 or 0.35 m).

Back-ups allowedRelaxed bound 0.15 m0.25 m0.35 m
034.7% inside limits, 30 of 49 rollouts hold38.1%, 27 hold49.7%, 16 hold, 29 crossings outside limits
153.7%, 23 hold87.8%, 7 hold100%, none hold
291.8%, 5 hold100%, none hold (configured)100%, none hold

Without a back-up the gate can still pivot to correct heading but cannot correct a lateral offset: the robot holds, or a wide relaxed bound lets it cross outside the limits. With at least one back-up, no crossing was outside the limits in any setting, because a robot released by the relaxed branch is at most the relaxed bound off at the check point and the last 0.7 m of tight tracking absorbs the rest. The 100% in the general response corresponds to the configured budget of two back-ups; with one back-up it is 87.8% with seven holds.

The general response covers the reversals of Section 5.4 and the robot's record of versions (Figure 5). We sincerely hope that Table 1, the factorial and these clarifications address the reviewer's concerns, and that the reviewer would consider raising the score. Thank you again for the detailed review; we wish you every success in your research and would welcome further exchange.

AuthorsOpenReview ↗

Thank you, and where each of your questions is answered

Dear Reviewer 47rh,

Thank you for the time and effort you have devoted to our paper, and for your detailed questions, which led to concrete changes. As the discussion closes on November 18 (AoE), we would be grateful to know whether our answers address them. For convenience, here is where each one is answered in the revised PDF.

Q1. How the stored conditions change the planning decision. Table 1 lists the runtime decision that each stored field changes, and Section 4.2 now states the two candidate-generation rules: where the stored route permits no detour, the only candidate is the stored line, and when the next goal lies within the planning horizon, every candidate returns to the line within the last metre before it. The comparison you proposed, the same centerline and executor with conventional tangent tracking and gait scheduling, is the first arm of the new factorial: after a late disturbance it brings 22.4% of ledge crossings inside the entry limits, against 100% with the stored events (88.6% under localization noise; Section 5.2, Tables 2 and 3).

Q2. Preparation, and the benefit over planning from a map. One teaching session, 27 minutes for the whole 271 m route, with nothing trained (Section 5.1, Table 6). Planning from the map and the goal sequence found a path for 13 of 29 segments.

Q3. Changes that invalidate a stored prior. Live obstacle sensing stays active on 58% of the route, where the planner detours around obstacles that are not on the map (Section 4, Appendix L). The exception branches are now measured (Table 5): with the configured two back-ups every crossing is inside the limits and none ends in a hold, and with at least one back-up no crossing is outside the limits.

Generality beyond one course and controller. The effect holds with our learned policies in contact simulation: after a push before the stairs, keeping the stored direction gives no fall in 180 climbs, against 9 falls without it (p = 0.0035; Appendix N). On the robot, the final version succeeded in all six competition runs (Section 5.5).

If our answers have resolved your questions, we would be sincerely grateful if you would consider raising your score. We would be glad to answer any remaining question before the deadline.

Comments on the paper

8 comments
AuthorsOpenReview ↗

General response to all reviewers

We sincerely thank the three reviewers for the time and care they gave to this paper; your questions led directly to the new experiments and to a clearer paper. All three reviews ask the same question: with the taught path and conventional route following held fixed, what do the stored passage events add? The previous version did not isolate this, and the revision now does. New text is in blue; details are in the individual responses.

The problem is a general one. A robot must follow a long outdoor route across stairs and ledges, each of which has to be entered within about ten centimetres and ten degrees, while localization and maps are imperfect and the locomotion skills are black boxes, since few teams can develop gaits or hardware for every site. In the language of options each skill has an initiation set, and for a black box it is unknown. When we built the system we found no open-source method that handled these conditions together. Our answer is simple and low-cost: one teaching session records how each transition was crossed, and the navigation layer enforces that entry after every local replan, with nothing trained for the site. We expect it to serve wherever skills come as they are, from field inspection and industrial sites to search and rescue. We also train locomotion policies with reinforcement learning ourselves, and they are the skills in our simulation experiments.

1. What the stored events add over tracking the same path (Section 5.2). This is the experiment the reviews asked for: path, gait schedule, sensing and feedback model are identical in every arm, and the first column is the plain baseline. Share of disturbed ledge approaches that cross inside the entry limits (10 cm, 10°), in kinematic simulation on the field reference:

Disturbance arrivesConventional tracking + gait schedule+ stored approach window+ entry gate
2.4 m before the ledge95.9%100%100%
1.2 m46.3%100%98.6%
0.76 m22.4%47.6%100%
0.76 m, with localization noise24.1%46.0%88.6%

With distance to spare the methods look alike. The difference appears where the problem is hardest, when a disturbance arrives suddenly and late, as after a detour that rejoins just before a ledge. There, under localization noise (3 cm, 1°), conventional tracking brings 24.1% of approaches inside the limits and the entry gate 88.6%; the gate cuts crossings outside the limits from 54% to 0.3%, and the remaining runs end in a safe hold. Without noise the gate refuses every approach outside the limits, the robot backs up and approaches again, and all rollouts end in a crossing. The gate costs time (4.2 s to 11.3 s on average from disturbance to crossing). A learned skill in contact simulation shows that entry matters physically: our PPO stairs policy passes 98.7% of entries within 5° of heading error and 32.7% beyond 25° (Appendix N).

2. Rejoining returns a displaced robot to undisturbed performance, and quickly (Section 5.3; contact simulation with our learned policies, 23 seeds). After a 1.2 m push, all 226 runs with our rejoining strategy return to the route and reach the next goal (2 of 226 when the robot holds), a median of 1.6 s later than without a push. Scored by the competition's waypoint rule (the body touches the 0.25 m disk of every goal in order, and a fall voids the run), they complete 75.7% of whole segments, against 0.4% with a hold and 77.4% for the same sets with no push; all but one of the remaining losses are falls on later terrain. The previous version used a stricter rule (centre within 0.20 m, a target 36% smaller in area), under which the three figures are 57.1%, 0% and 54.0%; the paper reports both. Against a plain baseline, the tracker's default recovery behaviour (turn and side-step; 61.4%), our strategy (walk forward while steering to the stored line) completes 75.9%, which is 14.5 points or 24% more.

3. On the robot (Section 5.5, Figure 5). The environment has 28 segments over more than ten kinds of terrain. In the logged end-to-end run (29 of 29 waypoints, 386.6 s), the gate refused the first approach to the first ledge, 0.16 m off the stored line; the robot backed up once and crossed within 0.07 m. Earlier versions, which stored the ledge as an ordinary waypoint, stalled there in both full-speed runs. On the competition day the final version completed six of six runs with factory gaits and no training for the environment. It won third place at the competition, which drew 2,999 submissions, and was the only finalist in its task to complete the route fully autonomously.

4. The reversals in Section 5.4. The reversals noted by Reviewers A6o1 and 47rh are between route-preparation variants that all carry the same passage fields, so they do not bear on the question above.

Thank you again for the thoughtful reviews. We hope this answers the shared question and would be glad to continue the discussion.

AuthorsOpenReview ↗

Response to Reviewer A6o1

We sincerely thank the reviewer for a precise and constructive review; the five suggested revisions shaped most of the new material. Their status:

Suggested revisionIn the revision
1. Isolate the passage fields (Q1)Done: new factorial, Section 5.2
2. Validate the complete pipelineThree additions (item 2)
3. Uncertainty of the entry-conditioning effectDone: 23 seeds, with intervals
4. Benchmark assetsDone: re-runnable package in the supplement
5. Physical comparisons across courses (Q2)Not possible at present (item 2); robot's record of versions added

Numbers refer to the revised PDF.

1. Isolate the passage fields while holding geometry and sensing fixed (Q1). Done in Section 5.2 (kinematic simulation on the field reference): centerline, gait schedule, observation input and feedback model are identical in all 16 arms of a 2^4 factorial, with one switch per mechanism of the previous version's factor. With a disturbance 0.76 m before a ledge, taught geometry with conventional route following brings 22.4% of crossings inside the entry limits, the stored approach window 47.6%, and the entry gate 100%; the gate's main effect is +64.5 points. So, with identical geometry and sensing, both event-anchored fields improve on conventional tracking, and the gate accounts for the larger part.

2. Validate the complete compiled-reference pipeline. Three additions narrow the gap between the stacks.

  • The factorial is repeated on the reference that the robot executed: 46.9%, 55.8% and 100%, with no hold.
  • Our two learned policies drive every contact-simulation experiment, including a full-route run (Figure 17; video in the supplement), and show the paper's premise: the stairs policy passes 98.7% of entries within 5° of heading error and 32.7% beyond 25° (Appendix N).
  • The two stacks share the navigation library byte for byte (35 source files) and 106 of 116 runtime values (Appendix B).

For background, the work began as a competition entry. Before the competition we trained our own locomotion policies with PPO (a walking policy for fast travel on gentle ground and a proprioceptive policy for stairs, rubble and slopes) and built the navigation stack on them in contact simulation; they also ran on the robot during development. Late in the preparation period the organizers opened the robot's factory-gait interface, which our final runs used. It is a paid service that is no longer available to us after the competition, which is why the simulation experiments use our own policies; these include no ledge skill, so the gate is measured in kinematic simulation. As suggested, related work now states the distinction from Lee et al. (2024), who integrate learned navigation and learned locomotion through hierarchical RL; our interface sits on a frozen library and trains nothing.

3. Uncertainty of the entry-conditioning effect. We added 20 seeds, and a re-run of the 192 earlier executions reproduced every outcome. The factor, now called the gait-entry lead, raises completion on undisturbed runs by 8.2 points under the rule of the previous version (95% interval 3.0 to 13.3) and by 6.2 points (−0.5 to 13.0) under the competition's waypoint rule; after a displacement it shows no effect (−4.4 and +0.9 points). The recovery result itself holds on 23 seeds: after a 1.2 m displacement all 226 rejoining runs reach the next goal, and whole-segment completion is 75.7% under the competition's rule, against 0.4% with a hold and 77.4% for the same sets undisplaced (57.1%, 0% and 54.0% under the stricter rule).

4. Benchmark assets. The supplement now contains an anonymized, re-runnable package: both frozen policy files, the reconstructed scene, the runner, drivers, analysis scripts and every per-run table. On 8 cells, the cleaned package reproduces the original trajectories byte for byte.

5. Physical comparisons across courses (Q2). New physical trials are not possible at present, because the factory-gait interface is no longer available to us (item 2), and they remain future work. What the revision adds is the robot's own record of consecutive versions on the same route (Figure 5, Appendix I). Versions that stored the ledges as ordinary waypoints stalled at the first ledge in both full-speed runs. The final version stores them as events: its logged run reached 29 of 29 waypoints in 386.6 s, after the gate refused one approach 0.16 m off the line and the robot backed up and crossed within 0.07 m, and on the competition day it completed six of six runs. On the stress cohort, all four arms of Section 5.4 carry the same passage fields, so that reversal is between variants of route preparation and sensing.

We sincerely hope the revision addresses the points raised and that the reviewer would consider raising the score. Thank you again for the careful review; we wish you every success in your research and would welcome further exchange.

AuthorsOpenReview ↗

Summary of changes in the revised PDF

New and rewritten text is in blue. The method is that of the previous version: the title and the algorithm of Section 4 are unchanged, and the earlier experiments are kept. The revision adds the experiments the reviews asked for and rewrites the text around them.

New experiments and evidence

  • Section 5.2, Tables 2-5, Appendix A: a kinematic factorial that separates the four mechanisms of the previous version's on/off factor (784 rollouts), with a disturbance-distance sweep, localization noise, and a sweep of the retry count and the relaxed bound.
  • Section 5.3, Tables 13-17, Figure 14, Appendix L: rejoining on 23 seeds (1,280 further executions), a third arm with the tracker's default recovery behaviour, a sweep of the displacement size (2,760 further executions), and two sweeps with an inserted box (2,576 further executions). Section 5.3 reports whole-segment completion under the competition's waypoint rule (the body touches the 0.25 m disk of every goal in order) next to the stricter rule of the previous version (centre within 0.20 m), with the undisplaced control, and states where the runs that do not complete are lost; all rejoining runs reach the next goal after the displacement, at a median cost of 1.6 s. Table 17 (Appendix L) gives both readings.
  • Appendix N, Figures 16-17, Tables 18-19: our learned locomotion policies in contact simulation, with a controlled entry-error sweep of the stairs policy (378 simulations), the archived end-to-end runs, and a map of one full-route run.
  • Section 5.5, Figure 5, Table 11, Appendix I: the robot's logs of the end-to-end run and of earlier versions on the same route.
  • Table 6, Appendix B: what preparing a route takes, and results of planning the route from the map alone.

Method (Section 4)

  • The entry gate, described in prose in the previous version, is written as Eqs. (7) and (8).
  • Table 1 lists each stored field and the runtime decision it changes.

Text

  • Abstract, introduction and discussion are rewritten around the isolating experiment; Figure 1 shows where runs end after a displacement next to their undisplaced control, and Figure 3 is redrawn.
  • Section 5.2 of the previous version is now Section 5.4, and its Section 5.4 is now Section 5.5; Section 5.2 is new.
  • Related work adds the literature on skill initiation sets and on composing pretrained or black-box skills, and the distinction from Lee et al. (2024).

Supplementary material

  • An anonymized, re-runnable benchmark package: the two frozen policies, the reconstructed scene, the simulation runner, experiment drivers, analysis scripts, per-run result tables, and a simulation video of the full route.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.