acceptodds
Under review as a conference paper at ICLR 2027

Beyond the Remembered World: Predictive 4D Belief for Persistent Navigation in Evolving Worlds

Abstract

Recent embodied navigation systems increasingly use persistent spatial memories and vision–language models to act across repeated visits to familiar environments. Persistent maps make prior observations available for subsequent decisions and reduce the need to reconstruct familiar scenes for every task. However, remembered observations need not describe the current world: objects may move outside the agent's view, even while it travels toward a predicted destination. Despite advances in memory retrieval and state prediction, existing methods do not jointly account for continued hidden world evolution during navigation and visibility-conditioned belief revision. We study navigation under hidden world evolution, where an agent must infer the latent current state from irregular histories and revise its decisions as time and evidence accumulate. We present , an embodied agent that turns timestamped 3D entity histories into a structured belief over whether the last observed state persists or the target has relocated. A frozen, zero-shot vision–language controller acts on this belief, while an event-driven filter forecasts states at candidate arrival times, incorporates newly informative, visibility-aware RGB-D evidence, and replans as the world evolves. We further introduce , a human-trace-grounded benchmark with scenes and tasks for evaluating state prediction and embodied decisions under controlled changes before and during navigation. Across simulation benchmarks and physical robot trials, improves navigation success and search efficiency relative to evaluated baselines. Paired controls show that temporal prediction is most useful when changes exhibit learnable regularity; ablations demonstrate the value of preserving uncertainty and updating beliefs from visual evidence.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied AI, Dynamic Object Navigation, Predictive 4D Memory, Long-Term Spatial Memory, Belief-Guided Navigation, Active Perception

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_eQJ7OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposed EvolvingNav, an navigation model which focuses on moving targets. The paper predicts the expected position of the moving targets, and deploys the navigation strategy accordingly.

Critical strengths

The research topic is interesting and important. The experiment is sufficient and comprehensive.

Critical weaknesses

The idea of this paper is interesting, and the experiments are fairly comprehensive. However, I think the performance gain of the proposed model could be analyzed more. EvolvingNav combines a structured temporal predictor, an arrival-time forecasting, a belief propagation, etc. The current ablations mainly remove broader modules and do not isolate these designs. More ablation studies that verify the effectiveness of these temporal components to be included would be better. For example, it would be useful to compare arrival time forecasting with only current-state prediction, learned belief propagation with a simple Markov model, and the full method with a version that does not reopen previously inspected locations.

Besides, most results are reported on EvoWorld-Bench, where object movements are generated using predefined household rules. Moreover, the temporal split uses different days from the same household histories for training and testing. Therefore, the model may learn patterns specific to the benchmark generator or individual households. Evaluation on unseen households or different transition rules would provide stronger evidence for generalization. (Though I haven't check whether there is more moving-object navigation benchmark)

Questions for authors

See weakness.

Reviewer_SW5iOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates lifelong robot navigation in dynamically changing environments, focusing on the challenge of historical spatial memories becoming outdated as object locations change over time. The authors propose EVOLVINGNAV, a framework that constructs a predictive 4D belief representation from timestamped 3D object histories. The method predicts target locations through a persistence–transition probability decomposition and integrates estimated arrival times at candidate destinations, visibility-aware Bayesian updates, and event-driven replanning to support target search and online recovery in dynamic environments. In addition, the authors introduce EVOWORLD-BENCH, a benchmark comprising 54 scenes and approximately 800,000 task instances. The effectiveness of the proposed approach is validated through simulation, cross-benchmark evaluations, and real-world robotic experiments.

Critical strengths
  1. The research problem is practically meaningful and relatively novel. This paper addresses the challenge of outdated historical memories in long-term robot operation within dynamic environments. It considers not only object movements occurring before a navigation query but also hidden state changes that continue during navigation execution. Compared with conventional static spatial memory and object navigation settings, this problem formulation better reflects the requirements of long-term robot autonomy and establishes a meaningful direction for predictive embodied memory.

  2. The method is well designed and closely aligned with the task requirements. The authors integrate time-conditioned state prediction, persistence–transition probability decomposition, arrival-time-aware target selection, visibility-conditioned negative observation updates, and event-driven replanning into a unified closed-loop framework. In particular, preserving probability mass for unknown locations, avoiding repeated use of correlated observational evidence, and allowing previously visited locations to be reconsidered are reasonable design choices for handling uncertainty in partially observable and continuously changing environments.

  3. The proposed benchmark represents a valuable contribution. EVOWORLD-BENCH includes 54 scenes and approximately 800,000 task instances, with object transition processes constructed using human activity records to capture temporal regularities. In particular, the paired design of static, regularly changing, and randomly changing environments helps distinguish the benefits of learning temporal patterns from those of general location statistics. This benchmark provides a useful evaluation resource for future research on lifelong navigation in dynamic environments.

  4. The experimental evaluation is extensive and provides convincing evidence. The authors conduct comparisons across multiple navigation benchmarks and specifically investigate scenarios involving target movement during execution, online search recovery, and cross-scene generalization. EVOLVINGNAV achieves notable improvements in first-target-check success rate, final search success rate, and path efficiency. In particular, the experiments involving continued target movement during navigation demonstrate stronger recovery capabilities compared with existing predictive navigation methods.

  5. The ablation studies are relatively comprehensive and effectively support the main contributions. The paper examines the effects of historical information, structured probabilistic prediction, full belief distributions, online evidence updates, and visibility calibration. In particular, the substantial decrease in search success rate after removing online evidence updates, together with the performance differences between regularly changing and randomly changing environments, helps clarify the sources of the method's improvements and the conditions under which it is most effective.

  6. Real-world robotic experiments further strengthen the completeness of the work. The authors validate the proposed method on different robotic platforms, including LYNX M20, X30, and Lite3, and provide results for indoor and outdoor target search as well as cross-platform transfer. These real-world experiments demonstrate that the proposed predictive memory and navigation mechanisms are not only effective in simulation but also show potential for practical deployment.

Critical weaknesses
  1. The independent contributions of certain key mechanisms could be further clarified. Although the paper provides relatively comprehensive module-level ablations, arrival-time-aware prediction and the dynamic reopening of candidate locations are important design choices that distinguish the proposed framework from existing predictive navigation approaches. However, direct ablations isolating these two mechanisms are still lacking. Consequently, their individual contributions to dynamic search and online recovery performance remain insufficiently quantified.

  2. Generalization to more complex behavioral changes remains somewhat limited. Although EVOWORLD-BENCH incorporates real human activity data and considers multiple types of dynamic changes, object transitions are still primarily generated through simulated mechanisms. The experiments indicate that the proposed method performs better in environments with learnable temporal regularities, while its advantages become less pronounced under random changes. Therefore, its applicability to unexpected or irregular object movements in real-world environments warrants further discussion.

Questions for authors
  1. Could the authors provide additional ablation studies to isolate the contributions of arrival-time-aware prediction and the dynamic reopening of candidate locations? For example, comparing target selection based on current-time probabilities versus estimated arrival-time probabilities, as well as evaluating online recovery performance with and without candidate reopening, would help clarify the individual benefits of these mechanisms.

  2. How reliable are the predicted beliefs when objects undergo unexpected movements that deviate from previously learned temporal patterns, or when long-term behavioral patterns change? Could the authors further discuss how the system adapts to such distribution shifts?

Reviewer_MuzGOpenReview ↗
Rating
4(+2)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work presents EvolvingNav and a benchmark EvoWorld-Bench. EvolvingNav is a zero-shot VLM-based navigation agent designed for changing indoor environments specifically objects. The new benchmark has 54 scenes across 162 dynamic states. The main contribution is a predictive 4D belief representation that separates permanent room geometry from movable objects, and updating where objects might have moved when they are no longer at their remembered spot. They evaluate this on their own EvoWorld-Bench, three existing benchmarks, and a real wheeled robot.

Critical strengths
  1. Realistic problem, since separating static room structure from movable objects seems like an intuitive way to tackle a very practical real-world problem.
  2. Also creates an accompanied Benchmark.
Critical weaknesses

Baselines may not be correct especially the re-implemented ones.

Questions for authors
  1. Why can this not beat findingdory that is doing better on findingdory dataset
  2. Why are some of the re-implementation numbers very low?
  3. How does the VLM latency affect real-world navigation time (and the execution times)
1 comment

Thank you for your thorough and transparent responses to Q1-Q3. My main concerns are addressed and the VLM latency tests show that live deployment is possible while the added ablations and other additions make the paper much stronger. I am raising my score to accept

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.