acceptodds
Under review as a conference paper at ICLR 2027

Transported-belief Planning for Multi-agent Search and Rescue under Uncertain Drift

Abstract

Maritime search for a drifting survivor requires allocating scarce assets while wind and ocean currents continually reshape where the survivor may be. Each route assignment affects both immediate detection and the team’s options for subsequent search. We develop a multi-agent planner that forecasts survivor movement, updates the shared forecast after each non-detection, and coordinates routes by anticipating future search opportunities and teammates’ movements. Our distributionally robust formulation derives detection bounds from the mean and covariance of survivor distributions transported by the forecast flow, capturing how routes share uncertainty through survivor drift—Lagrangian transport—rather than through geographic proximity alone. Across 180 simulated search settings based on OSCAR current data from five ocean regions, the planner achieves 50.1% expected mission success, outperforming common benchmark approaches including classical lawnmower search and reinforcement-learning policies. Ablation studies further show that forward planning and teammate modeling both contribute to performance, with gains sustained across the tested fleet sizes, mission scales, and environment conditions. Our work suggests the importance of coordinating current assignments around predicted survivor movement while preserving valuable opportunities for the team’s next search.

Primary area:
optimization
Keywords:
search-and-rescue, multi-agent planning, distributionally robust optimization, Gaussian processes, Lagrangian uncertainty

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_LrqVOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents transported-belief rollout (TBR) a framework to plan search and rescue opertations for maritime suvivors, meaning drifting persons at sea. Challenges are coordinating multiple assets under unknown dirft conditions and optimizing fast detection.A Gaussian process is fitted to observed ocean currents in which survivor particles are transported through sampled current fields. This procuduces a shared trajectory-level belief that is updated after each unseccessful search. Assets are coordinated via receding-horizon rollout that scores the joint detection of teammates' routes and the search options preserved after a miss. The framework is evaluated on 180 seach configurations spanning five OSCAR, a ocean current dataset, regions against classical search, coordinator heuristics and learned RL policies, where it improves against PPO by ~3.6pp. Ablations show the contributing factor of each component of the framework.

Critical strengths

None

Critical weaknesses
  • While the application scenario is interesting and important, the paper teaches little about representation learning. The only learned component is a standard two-output GP for the current field. The contribution therefore seems to lie mainly in planning/operations research, and it is unclear whether ICLR is the best venue.

  • The PPO baseline (46.5%) is not far from the more sophisticated TBR framework (50.1%). The other multi-agent RL baselines were trained for only 500–2,000 episodes with default hyperparameters. A comparison with model-based RL (e.g., DreamerV3) or learned-value search (e.g., MuZero-style) would be more informative and might close the gap.

  • Clarity on the problem description and the RL setting could be improved. A formal problem statement (e.g., explicit state, action, observation, and reward) is missing; notation is dense; and the "manager" framing makes the main text hard to follow.

Questions for authors
  • How does TBR compare against model-based RL or MCTS with learned value functions, and against PPO trained with a substantially larger budget (e.g., until convergence)?
Reviewer_kLmwOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper focuses on a multi-agent approach to maritime search and rescue under uncertainty. The paper includes many theoretical aspects that lead to the algorithmic approach presented. An evaluation shows improvements on top of a range of baseline methods.

Critical strengths

Results show improvements on metrics such as the success rate in comparison to a range of baselines that include learned RL policies and classical search.

Critical weaknesses

I find the paper to be difficult to follow at the level of technical details. I expect that many readers will have a hard time reading the paper and I consider that a major rewriting is necessary. For example, concepts that appear to be important are used before being defined, which has the effect of making the paper more difficult to follow. Such concept include path identity, shared exposure / drift exposure, downside (downside protection, downside analysis etc.). It is hard to understand statements such as the following one, for example: “path identities reveal repeated or complementary coverage within a scenario, while changes in team detection across scenarios reveal shared drift exposure.” Despite looking backwards in the paper to identify definitions of concepts mentioned in this sentence, I find it to remain hard to understand and confusing. This example is representative for the overall feeling I had with reading the paper. Adding to the main paper clear definitions and a clear and simple example that walks through the steps of the approach would help, as part of a more comprehensive effort to improve the readability of the main paper from the beginning to the end and make it truly accessible to an audience.

Questions for authors

No specific questions.

Reviewer_9jYiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies multi-agent maritime search and rescue, where ocean currents and wind make it difficult to predict the location of a drifting survivor. The authors use a Gaussian process fitted to OSCAR current data to generate possible drift scenarios and survivor trajectories. When a search is unsuccessful, the probabilities of these trajectories and scenarios are updated using Bayes' rule.

The proposed method, Transported-Belief Rollout (TBR), plans search routes for multiple assets while considering what other team members are doing. It evaluates possible routes over several steps, taking into account the probability of detecting the survivor and the future movements of teammates. Only the first part of the selected route is executed before replanning. The paper also presents theoretical results explaining why keeping complete trajectory information is useful for sequential search, along with a moment-based lower bound for detection under uncertainty.

The authors evaluate TBR across 180 simulated configurations using ocean-current data from five regions. TBR achieves 50.14% mission success, compared with 42.59% for the transported-belief coordinator and 46.51% for PPO using the same belief information and action menu. Interestingly, a simpler greedy rollout achieves similar performance (50.27%) with lower planning cost, while deeper UCT search reaches 51.50%. The ablation studies suggest that lookahead, teammate coordination, and trajectory-level information contribute to performance. Additional experiments examine different fleet sizes, environmental conditions, and computational costs.

Critical strengths

1.Experimental evaluation The paper evaluates the method across 180 configurations, with paired comparisons and confidence intervals. TBR improves both mission success and detection time compared with the transported-belief coordinator. The improvements also hold across the tested ocean regions, different fleet sizes, and wind conditions. The authors provide useful baseline comparisons, including PPO with the same belief information and action menu. The paper also reports cases where simpler or more expensive alternatives perform better. For example, greedy rollout achieves similar success (50.27% vs. 50.14%) with lower computational cost. 2. Clear evidence for trajectory-level coordination. The paper shows that keeping track of which survivor trajectories are covered by teammates is important for coordinated search. Removing this shared trajectory information reduces mission success by 10.67 percentage points. Together with Proposition 1, this provides useful evidence that trajectory-level information matters when planning searches for multiple agents.

Critical weaknesses
  1. The novelty of the proposed planning method is not clear . The paper combines trajectory-based beliefs, Bayesian updates, multi-step rollout, and teammate coordination. While the combination is useful, several of these ideas have already been explored in moving-target search and multi-agent planning. For example, the authors acknowledge that SAROPS transports survivor trajectories and updates their probabilities after unsuccessful searches. I would have liked a clearer discussion of how TBR differs from earlier work on constrained-path moving-target search (Eagle & Yee, 1990; Royset & Sato, 2010) and multi-robot search (Hollinger et al., 2009).

The experimental results also make me question the advantage of the specific TBR planning procedure. Greedy base-policy rollout achieves 50.27% success compared with TBR's 50.14%, while taking considerably less computation time (2.62 versus 10.00 seconds in the separate timing study). Deeper UCT search achieves 51.50%, although it takes longer and has slightly lower mid-mission detection.

I agree that keeping trajectory-level information is useful for coordinating multiple agents, and the ablation results support this. However, I am still not convinced what the proposed TBR algorithm adds beyond existing rollout approaches. This is my main concern with the paper, and a more direct comparison with related planning methods would help clarify its novelty.

  1. The contribution of the learned drift model and robustness formulation is not sufficiently demonstrated.

Although the authors use OSCAR ocean-current data, the simulation relies on averaged current fields and a prescribed wind model. The GP forecast also shows rather weak predictive performance, with spatial correlations between 0.00 and 0.21. Even when the planning ensemble is re-centered on the true current, TBR improves by only 1.35 percentage points in the reported experiment. This raises questions about how much the learned current model contributes and whether the results would hold under more realistic, time-varying ocean conditions.

I also have some concerns about the practical role of the distributionally robust formulation. The main TBR experiments use κ = 0, meaning the robustness penalty does not contribute to the headline performance improvement. The separate risk-sensitive experiment shows a relatively small improvement in lower-quarter success (0.82 percentage points), with a slight reduction in mean performance (0.22 points). The paper also does not establish whether the assumptions needed for the theoretical guarantee hold for the true drift distribution.

Overall, the experiments show that the approach works well in the proposed simulation environment. However, I would like to see stronger evidence for the contribution of the learned drift model and the robustness formulation, particularly under more realistic environmental uncertainty.

Questions for authors
  1. Greedy base-policy rollout achieves almost the same success as TBR (50.27% vs. 50.14%) with much less computation, while deeper UCT search performs slightly better. What is the main advantage of the proposed TBR approach over these alternatives? I would particularly like to understand what the three-leg planning procedure adds beyond existing rollout methods.
  2. I would like to understand how TBR differs from earlier moving-target search methods, such as Eagle & Yee (1990), Royset & Sato (2010), and the multi-robot search approach of Hollinger et al. (2009). Could the authors explain what is new compared with these methods? A direct experimental comparison with at least one relevant method would help clarify the contribution.
  3. The GP forecast shows relatively weak prediction accuracy in the reported experiments. How would TBR perform with time-varying ocean currents instead of the averaged current fields used here? I am also interested in how sensitive the results are to errors in the wind model beyond those already tested.
  4. The main TBR experiments use κ = 0, and the risk-sensitive version provides only a modest improvement in lower-quarter success. How important is the distributionally robust formulation to the proposed approach in practice? Is there evidence that the moment assumptions used in the theoretical analysis adequately capture uncertainty in the true drift distribution?
  5. The method assigns agents one at a time using a fixed order. Does changing this order affect the results? It would be useful to know whether the reported improvements remain consistent when agents are assigned in a different or random order.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.