acceptodds
Under review as a conference paper at ICLR 2027

Knowing More, Completing Less: Reliability Learning in Lifelong Robot Fleets

Abstract

A robot that stays still may have failed, or it may be waiting for another robot that failed. Learning from such motion is a statistical problem; using what is learned is a coordination problem. We study their connection in lifelong multi-agent path finding. Our central observation is a reversal: under the same planning rule, replacing fleet-average reliability with exact individual reliabilities increases completed tasks by 7.96% in a room but decreases them by 13.09% in a warehouse, with the corresponding sign in all six initial-state averages per setting. To understand this gap, we derive the position-feedback likelihood and characterize exactly which immediate task-value comparisons are identifiable. A task-value difference can be identifiable even when individual reliabilities, or both competing values, are not. We then develop a motion-censored release preference that learns from unambiguous actuator outcomes and refines equally preferred next moves, yielding a 9.76% gain in a two-room evaluation with a frozen method and previously unused seeds. A trajectory decomposition associates the room gain with fewer propagated stalls and the warehouse loss with a growing deficit in proposed moves. A four-robot construction then shows how better immediate task completion, more movement, and fewer robots in cells with at most two neighbors can all lead to fewer tasks after two steps. Together, these results connect what a fleet can learn from motion to the future its planner creates.

Primary area:
optimization
Keywords:
lifelong multi-agent path finding, multi-agent coordination, execution uncertainty, reliability-aware planning, topology-aware planning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 2 reviews, above 0% of the others.
Reviewer_KGyhOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents 3 main contributions. 1. The feedback model and identifiability theory: when can individual reliability (per robot) be inferred from joint movement and which plan values can be inferred when individual reliability cannot. 2. MCRP: Guided-PIBT over Beta estimates of each robot’s reliability. 3. The reversal: in some cases knowing the individual reliability is counterproductive.

Critical strengths
  1. The result that plan-value differences can be identified when individual reliabilities are not can be useful for learning from censored motion.
  2. Negative results are reported in full, and code and audit materials are provided.
Critical weaknesses
  1. Claims are tied to specific settings and wrapped in caveats (“descriptive, not causal,” “development, not confirmatory,” “does not isolate a mechanism”). It is unclear what a reader should conclude or do differently.
  2. The central reversal is not turned into guidance. The paper gives no rule for when per-robot reliability helps, no remedy for when it hurts, and no prediction for new settings. The authors also note that an optimal planner would never be hurt by the extra information, which further limits what the reversal implies.
  3. Negative results receive long explanatory detours e.g., “Why a better next step can create a worse future” and the warehouse appendices. These take space away from the main thread without producing a usable conclusion.
  4. The one general result, that differences between plan values can be identified when individual reliabilities cannot, is buried under cryptic headings such as “Paths expose labels; cycles expose products.”
Questions for authors
  1. There is a general sense (for me) that the writing relied heavily on AI tools. Was that the case? Please note that AI tends to produce text that suffers from weaknesses 1-4.
Reviewer_rpgTOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates lifelong multi-agent pathfinding (MAPF) under actuator failures, where the planner must learn each robot’s actuator reliability from position feedback. The authors first analyze the structure of this feedback, derive the corresponding likelihood function, and establish identifiability criteria for both the immediate value of a given plan and the task-value difference between alternative plans. They then propose a method called MCRP, which uses censored Beta counts to estimate each robot’s reliability. These estimates are used as a tie-breaking mechanism in the Guided-PIBT algorithm whenever multiple candidate grid cells have equal guidance distance. The authors report that MCRP improves completed tasks by 9.76% over Guided-PIBT on two room maps. In addition, under a fixed rule, exact individual reliabilities raise completed tasks by 7.96% on a room map but reduce them by 13.09% on a warehouse map, compared with the fleet mean.

Critical strengths
  1. This paper reveals a noteworthy phenomenon: under a fixed decision rule, providing exact information about individual reliability to the planner can either increase or decrease task completion, depending on the environment.
  2. The study is supported by a rigorously paired experimental design, detailed experimental protocols, and sound theoretical derivations.
Critical weaknesses
  1. The paper doesn‘t commit clearly to either proposing a method or revealing a phenomenon. In fact, the phenomenon highlighted in the title is exactly the scenario where the proposed method MCRP fails.
  2. The methodology contribution is incremental. The learning component is merely frequency estimation for individual robots, used only for tie-breaking. Ablation studies indicate that estimation accuracy does not dictate overall performance. Therefore, it is necessary to exclude the possibility that the observed gains come from perturbations introduced during tie-breaking. Furthermore, the theoretical framework presented in Section 3 is not actually used in the MCRP method.
  3. The locked evaluation is limited to only two room maps, without analysis of the MCRP method's broader applicability. Furthermore, the paper doesn't provide a solution for the failures observed in the warehouse map.
  4. Experimental evaluation is restricted to the backbone planner, the authors' own ablation variants, and the HR method. It lacks comparisons against highly relevant and competitive baselines. Additionally, the experimental settings used across experiments are inconsistent, and the computation of certain reported results, such as the 9.76% improvement, is not defined.
  5. The related work section lacks substantive comparisons with relevant existing methods, the research gap is not clearly defined.
  6. The presentation requires improvement. There are inconsistent terms throughout the text, Figure 5 contains formatting errors, and the appendix reads like experimental logs.
Questions for authors
  1. How is the 9.76% macro-average computed in the paper?
  2. Would MCRP’s gain persist if the reliability estimates are replaced with fixed random per-robot scores?
  3. Why does individual reliability information reduce proposed moves on the warehouse map?
  4. Why do Sections 5.1 and 5.2 adopt different seeds and fault-tape designs?
  5. Can the method scale with the fraction of fragile robots or the number of agents?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.