acceptodds
Under review as a conference paper at ICLR 2027

JaxNaviSim: A GPU-Accelerated, Data-Driven Simulator and Learning Benchmark for Maritime Navigation

Abstract

Coordinating vessels through the shared waterways of busy ports grows more challenging as vessel traffic increases. Port authorities need realistic simulations to assess how arrival schedules, anchorage assignments, and routes affect vessel throughput and navigation safety. We introduce JaxNaviSim, a data-driven simulator and learning benchmark coupling Automatic Identification System (AIS) data with fast, controllable traffic simulation. First, we release a dataset of  141 million AIS records from over 30K vessels across eight high-density regions of the Singapore Strait, among the world's busiest waterways, with  6.7 million vessel-hours. Unlike several existing AIS datasets which mostly cover open seas, ours captures close-quarters vessel traffic in the narrow strait. Second, we formulate navigation as a goal-reaching task, and implement a scalable JAX simulator supporting historical replay, selective vessel control with batched GPU execution. We also incorporate multiple vessel dynamics models and sea surface currents. Finally, we benchmark imitation learning (IL), offline and online RL, with ablations of key design choices. Learned policies reach high goal completion, but collisions persist. We therefore propose Predictive Action Refinement (PAR), which refines actions using short-horizon interaction forecasts, substantially reducing collisions. Our contributions enable research on scarce maritime resource allocation, traffic modeling, digital twins, and provide large real-world testbed for IL/RL methods.

Primary area:
datasets and benchmarks
Keywords:
maritime navigation, imitation learning, reinforcement learning, GPU-accelerated simulation, data-driven benchmarks, closed-loop evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_gk8vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces JaxNaviSim, a GPU-based maritime navigation simulator and benchmark built from real AIS records. It supports traffic replay and single- and multi-vessel control, and compares imitation learning and reinforcement learning methods. It also introduces PAR, which checks predicted collisions and adjusts proposed actions. Experiments across eight regions examine navigation performance, collision rates, and different dynamics settings.

Critical strengths

1: The platform addresses a useful problem. Combining real vessel trajectories with parallel simulation provides a useful resource for studying navigation in dense maritime traffic.

2: The experiments cover several learning approaches and settings. The comparisons show that better goal reaching does not necessarily mean fewer collisions. PAR substantially reduces collisions in the reported experiments while largely maintaining goal-reaching performance.

Critical weaknesses

1: The test set is used for model and parameter selection. Checkpoints and the PAR configuration are selected using test-set results. This makes the final evaluation less independent and may produce optimistic results. A separate validation set and an untouched test set are needed to establish the reliability of the reported gains.

2: PAR’s safety benefits and technical contribution need stronger support. The main results use simplified position and heading updates and interpolated AIS states. It is unclear whether the gains hold under stricter vessel dynamics or delayed, missing, and noisy observations. PAR also combines fairly standard ideas: constant-velocity prediction, collision checking, and candidate action selection. Comparisons with established safety-filtering methods would help establish what its specific design adds.

Questions for authors

1: How do the results change when checkpoints and PAR parameters are selected on a separate validation set and evaluated on an untouched test set? 2: Do PAR’s collision reductions hold under the alternative dynamics models, and when surrounding-vessel observations are delayed, missing, or inaccurate? 3: How does PAR compare with established collision-avoidance or safety-filtering methods under the same observations and motion constraints? 4: The main text relies heavily on text and tables, while the appendix diagrams are too schematic to explain the key mechanisms. Please add a clearer framework figure showing how AIS data become simulation scenarios, how policies control vessels, and where PAR intervenes. A concrete avoidance example showing the proposed action, alternatives, and final choice would also help.

Reviewer_DWy4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents JaxNaviSim, a data-driven simulator and benchmark for maritime navigation. Based on AIS vessel-trajectory recordings from the Singapore Strait, the authors derive goal-reaching navigation tasks in eight regions, covering two settings: anchorage groups and traffic separation scheme (TSS)/intersection regions. The JAX-based simulator supports batched GPU execution and is about three orders of magnitude faster than the publicly available, CPU-based implementation of the earlier work ShipNaviSim. The authors evaluate a range of baselines (open-loop trajectory prediction, imitation learning, offline RL, and online RL) and propose a test-time action-filtering method, Predictive Action Refinement (PAR), to improve collision avoidance. PAR substantially reduces the collision rate of the flow-matching behavior-cloning policy (FM-BC), from 20.4–45.1% to 1.1–3.4%, while largely preserving goal-reaching rates. The AIS data and the navigation tasks are made public.

Critical strengths
  1. Large-scale AIS dataset (~141 million records from over 30,000 vessels) and challenging close-quarters navigation tasks in a real, dense port environment.
  2. Fast simulator: batched GPU execution reaches millions of vessel steps per second on a single consumer-grade GPU, and the simulator supports historical replay, single-ego control, and multi-vessel self-play.
Critical weaknesses
  1. The presentation of results lacks clarity. There is no aggregation across tasks or regions, no single benchmark score that would provide an ordering of the approaches, no normalization (e.g., relative to the best method or to the expert/human replay, which by construction achieves R_G = 1 and zero collisions), and the best results are not highlighted in the tables. In addition, a detailed investigation of the expert data would be valuable.
  2. While the dataset is certainly a valuable resource, the other contributions may be of limited interest to the general ICLR audience: the methodological novelty lies mainly in PAR, a hand-designed test-time filter.
  3. AIS data alone could leave blind spots, since not all vessels (e.g., small craft) carry an AIS transmitter. This could lead to underestimated collision risks and should be discussed.
Questions for authors
  1. Do the authors plan to provide reactive background agents (e.g., rule-based or learned) so that replayed traffic can respond to the controlled vessel? Or would that be the case, if selecting all vessels controllable?
  2. Under which license is the MPA-provided AIS data redistributed, and how are vessel identifiers (IMO/MMSI) handled with respect to privacy?
  3. Could the authors discuss the relation of their work to the NAVSIM benchmark (https://arxiv.org/abs/2406.15349)?
  4. Would the authors consider presenting the results graphically rather than only in tables, e.g., using the aggregate plots from rliable (https://arxiv.org/abs/2108.13264)?
Reviewer_wgyyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

JaxNaviSim combines AIS-derived vessel trajectories from eight Singapore Strait regions with a batched JAX simulator and navigation benchmarks. Policies control one vessel among replayed traffic or multiple vessels jointly, using local observations and a destination inferred from the recorded passage.

Critical strengths
  • The resource brings together trajectory data, episode construction, selective vessel control, and accelerator-native execution in a maritime setting that is poorly served by existing learning benchmarks. The eight regions include anchorage, crossing, and traffic-lane settings, and the interface supports both replay-based and joint-policy evaluation. This is a useful contribution in its own right, without requiring a new learning algorithm or a fully calibrated hydrodynamic digital twin.

  • The behavioral evaluation exposes a consequential distinction between reaching a destination and avoiding collisions. Tables 3 and 4 report large PAR reductions in sampled collision rates across all four anchorage regions while retaining similar or better goal completion in most regions. The same selected FM-BC checkpoints are used with and without PAR, and the appendices provide per-seed results, controller details, and limitations. These observations are useful even though the current selection protocol prevents treating them as a fully independent test.

Critical weaknesses
  1. The Oct–Dec test split is used to choose models rather than only to evaluate them. Table 8 selects FM-BC checkpoints by test goal-reaching rate, with tracking error breaking ties. Table 15 applies test-based checkpoint selection across the learned baselines, and Section I says the reported PAR configuration was chosen from a test-set parameter study. This makes the future-traffic split part of method development. Running three seeds does not restore an independent evaluation, and the issue also affects comparisons with GLRP, whose checkpoint is selected on separate validation days according to Section E. The disclosure is helpful, but it does not remove selection bias or establish that the rankings survive validation-only selection. For a paper offering a reusable benchmark, this is a central protocol problem. Separate validation from final testing for checkpoint and controller choices, then re-establish the comparisons with those choices frozen. Until then, describe the current results as exploratory rather than independent held-out estimates.

  2. I cannot interpret the approximately 985-fold speed comparison as acceleration of the same simulation workload. Section M says the GPU runner omits rewards and full collision and tracking metrics, whereas the CPU reference computes built-in metrics. It also consumes only the final-time x-coordinate slices of observations, allowing other observation computations to be pruned by compilation. Whether this pruning actually occurs is not measured. Consequently, counting valid vessel transitions does not establish that a complete policy observation or benchmark step was produced. Calling this a microbenchmark is appropriate, but does not make the CPU/GPU ratio comparable. Waymax at NeurIPS 2023 separates transition timing from metric-inclusive stepping, illustrating a feasible distinction. Report a matched replay workload that consumes the required observations and uses common metrics and termination, and distinguish it from a learned-policy rollout measurement. Alternatively, retain the current numbers as limited microbenchmark results without presenting their ratio as a like-for-like speedup.

Questions for authors
  1. Are validation-selected or fixed-final-checkpoint evaluations already available for the main comparisons? Please also specify which test populations were consulted when choosing PAR configuration C3. Existing results or a precise selection history could materially change my assessment. I am not requesting a new training campaign during rebuttal.

  2. Is there an existing timing result that consumes complete observations and matches the CPU reference's metric work and termination rules? If not, which narrower computational claim do the authors intend the reported ratios to support? The distinction between a limited replay measurement and usable policy-rollout throughput is central to my second concern.

1 comment

I keep the original rate. The matched replay benchmark addresses my timing concern. My main concern remains unresolved because QC-FQL was trained with FM-BC teachers selected on the original Oct–Dec test split, and reselecting the saved QC-FQL checkpoints on Oct cannot undo the training-stage dependency. PAR also retains C3, which was chosen using the AEW and AESPA test populations, meaning the re-analysis still falls short of an independent Nov–Dec comparison for these results.

Reviewer_mjXyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a new dataset and benchmark for data-driven learning-based marine-time navigation. The paper also compares a wide variety of policies including offline and online RL based methods for the navigation policy over the benchmark. To minimize collision, the paper further proposes a test-time heuristic for action replacement named Predictive Action Refinement.

Critical strengths

The benchmark and dataset size is scalable with 141 million AIS records, 30K vessels and 6.7 million vessel hours. The comparison also involves multiple baseline methods across trajectory-prediction based, imitation learning based, online and offline RL methods.

Critical weaknesses

The technical novelty is limited. The proposed PAR action refinement largely is heuristic based. Also, there appear to be many hyper-parameters including future horizon lookahead, goal movement thresholds etc. So it's unclear how robust the method is with all these different parameters.

Questions for authors

It would be good to add the size and scale comparison of the proposed benchmark dataset with prior benchmarking works.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.