acceptodds
Withdrawn from ICLR 2027

PreAct-Nav: Agentic Reasoning Before Action for Urban Navigation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

18 views

Abstract

Urban navigation requires embodied agents to pursue long-horizon goals through local decisions based on egocentric observations. However, existing agentic navigation methods often struggle to translate distant goals into coherent local decisions in large-scale physical environments. Their reliance on linguistic reasoning over transient observations or limited history constrains anticipation of the consequences of actions and future conditions, despite the importance of such foresight for navigating long and complex urban routes. To bridge this gap, we propose PreAct-Nav, an agentic navigation framework that equips frozen policies with anticipatory reasoning for robust urban navigation. Our central idea is to anchor local decisions in persistent medium-horizon subgoals, assess the consequences of predicted actions before execution, and continually update the reasoning context using actual outcomes. At its core, a navigation memory module maintains the active subgoal and relevant experience across decisions, translating distant goals into actionable intermediate objectives. We further introduce a predictive world sandbox that uses an action-conditioned world model (AC-WM) to forecast world dynamics conditioned on candidate movements. A vision-language model (VLM) reasoner interprets these predictions under the current subgoal to retain or revise actions. After execution, real observations are used to assess outcomes, correct inconsistent assumptions, and update memory to continue or reformulate the subgoal. Extensive evaluations demonstrate that the proposed PreAct-Nav improves action selection through memory updates and visual prediction, with more pronounced gains on longer routes and routes with more turns.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Urban Navigation, Agentic Reasoning, Action-Conditioned Prediction

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_6NaiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PreAct-Nav, a training-free framework for urban navigation that integrates structured subgoal memory, action-conditioned video prediction, and VLM-based action review and revision. By combining predictive reasoning with execution feedback, the framework aims to improve action selection without updating the underlying model parameters. Experiments on UrbanNav and CityWalker, as well as evaluations with different navigation policies, demonstrate improvements in navigation performance.

Critical strengths
  • The paper is generally well written and clearly presented. It addresses a relevant problem in urban navigation and proposes a training-free framework integrating VLM reasoning, predictive world modeling, and execution feedback.

  • Experiments across multiple datasets and navigation policies show promising performance improvements.

Critical weaknesses
  • The technical novelty remains insufficiently established relative to existing closed-loop planning and memory-based methods. The contributions of the proposed reasoning and memory mechanisms are not clearly distinguished from generic VLM-based action refinement, while the reliability of prediction assessment and the computational cost of the framework are not sufficiently validated.
Questions for authors
  1. PreAct-Nav combines structured subgoals, action-conditioned video prediction, and execution feedback in a training-free framework. However, prior work such as CLOVER [1] has explored generative visual planning and closed-loop feedback, while Reflexion uses execution feedback to improve subsequent decisions. What fundamental technical advance does PreAct-Nav offer beyond these approaches, particularly in its use of VLM-based action review and revision?

  2. Table 4 shows that the Reasoning Harness alone improves SR by 2.42 percentage points. However, it remains unclear whether this gain stems from the proposed Preview–Review–Revision mechanism or from generic VLM-based action critique and refinement. A comparison with a simpler critique-and-refine baseline using the same VLM and a comparable number of calls would help establish the specific contribution of the proposed reasoning design.

  3. PreAct-Nav maintains intermediate subgoals through four fields: Objective, Visual cues, Open question, and Status. Although the ablation results show improvements from Navigation Memory, it remains unclear whether these gains arise from the structured subgoal representation itself or simply from providing additional interaction history. Would simpler history-based context or alternative memory designs achieve comparable gains?

  4. Table 4 shows that adding the Predictive Sandbox to Navigation Memory improves SR by 1.01 percentage points. Although Figure 3 suggests greater benefits on longer routes and routes with more turns, Tables 12–13 in Appendix G report additional video generation and VLM inference costs. What is the end-to-end latency of the complete system, and how does the actual frequency of predictive rollouts affect the trade-off between performance and computational cost?

  5. PreAct-Nav relies on the same VLM to assess predicted videos, evaluate candidate actions, and update memory based on execution feedback. However, inaccurate predictions may mislead the reasoner, and some feedback labels in the appendix are generated by the VLM itself rather than independent ground-truth annotations. Could the authors provide an independent evaluation of prediction quality or analyze the relationship between the reasoner's judgments and actual navigation outcomes to better establish the reliability of the proposed closed-loop mechanism?

[1] CLOVER: Closed-Loop Visuomotor Control with Generative Expectation for Robotic Manipulation, NeurIPS 2024.

Reviewer_7FhzOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PreAct-Nav, a training-free agentic navigation framework that integrates navigation memory, action-conditioned world models, and VLM-based action review. The method maintains intermediate navigation subgoals, predicts future visual observations using a pretrained world model, and revises proposed actions before execution. Experiments on UrbanNav and CityWalker demonstrate improvements across different navigation policies.

Critical strengths
  1. The paper addresses an important problem in long-horizon urban navigation by incorporating anticipatory reasoning and execution feedback.
  2. The framework achieves consistent improvements across multiple navigation policies without additional model training.
  3. The authors provide component ablations, route-complexity analyses, and a qualitative real-world demonstration.
Critical weaknesses

W1. Limited methodological novelty. My primary concern is that the proposed method is largely an agentic harness integrating existing pretrained models through structured prompts and tool calls, rather than introducing a substantially new learning algorithm or model architecture. Navigation Memory mainly maintains intermediate subgoals and historical feedback through VLM-based state updates, while the Predictive Sandbox directly employs an existing action-conditioned world model. The preview-review-revise workflow also follows established agentic reasoning paradigms. Although training-free integration is valuable, the paper does not sufficiently demonstrate what fundamentally new methodological insights this framework provides beyond existing memory-augmented reasoning and world-model-based planning approaches.

W2. Insufficient validation of world-model predictions and limited incremental benefits. The effectiveness of the proposed framework relies on the assumption that generated future observations provide reliable evidence for navigation decisions. However, the paper lacks systematic quantitative evaluation of prediction accuracy, geometric consistency, or correspondence between predicted and actual future observations. Furthermore, adding the Predictive Sandbox to the reasoning harness with Navigation Memory improves success rate only from 82.37% to 83.38% on UrbanNav test-unseen.This relatively modest improvement raises questions about whether the additional world-model reasoning provides sufficient benefits over simpler action-review strategies.

W3. Substantial inference overhead. The proposed method introduces considerable computational overhead. According to Appendix G, a single world-model rollout requires approximately 11.37 seconds on an H200 GPU, excluding additional VLM reasoning calls. The paper does not adequately establish whether these computational costs are justified by the performance improvements or whether the framework can support practical real-time navigation.

W4. Insufficient quantitative real-world validation. Although Appendix B presents a real-robot navigation demonstration, it is limited to a qualitative example. No quantitative real-world success rates, collision statistics, or systematic comparisons with baseline policies are reported. This limits the evidence for robustness and deployment feasibility in dynamic urban environments.

Questions for authors
  1. What are the fundamental methodological contributions beyond combining existing agentic reasoning, memory, and world-model components?
  2. How accurately does the world model predict navigation-relevant future observations, particularly around obstacles and occluded regions?
  3. Can the authors compare against simpler action-review methods under matched inference budgets?
  4. Can the authors provide quantitative real-world evaluations and end-to-end decision latency measurements?
Reviewer_tRiyOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PreAct-Nav a training-free framework that improves frozen navigation policies through action review and revision. The framework combines navigation memory, which maintains recent experience and persistent intermediate goals, with an action-conditioned world model that predicts future visual observations. A VLM uses these predictions to assess proposed actions and revise them when necessary, while observations obtained after execution update the memory and subgoal. Experiments on UrbanNav and CityWalker report improvements across navigation policies and VLM reasoners.

Critical strengths

The method is well motivated; it combines persistent medium-horizon subgoals with action-conditioned visual prediction to guide local navigation decisions. Predicted observations inform action review and revision, while actual execution feedback updates navigation memory and subgoal progress.

Critical weaknesses
  1. My main concern is the evaluation setup. UrbanNav and CityWalker use recorded trajectories, but PreAct-Nav can generate revised actions that depart from those trajectories. The paper does not explain how these actions determine the agent's next position and visual observation, particularly when the proposed movement deviates from the recorded path. The original benchmarks do not directly establish the closed-loop evaluation procedure used here. The authors should clarify how observations are obtained, how off-route movements are handled, and how navigation success is determined. Without these details, it is difficult to assess the validity of improvements. (Happy to increase score if addressed)

  2. The memory ablation removes both historical observations and persistent subgoals simultaneously. It is therefore unclear whether the improvements come from maintaining subgoals or simply providing additional historical context.

  3. Important related works missing: [1] NavForesee: A Unified Vision-Language World Model for Hierarchical Planning and Dual-Horizon Navigation Prediction (CVPR 2026) [2] Can MLLMs Find Their Way in a City? Exploring Emergent Navigation from Web-Scale Knowledge (EACL 2026) The distinction of the paper's method from these works is important.

Questions for authors

I am curious what happens when navigation routes become much longer, particularly at the city scale. Does world modeling still help in such cases? My concern is that hallucinated future predictions might accumulate over longer trajectories and hurt navigation more than they help. In such settings, language-based methods that generate future routes, as explored in prior works[2], might perform better. It would be interesting to evaluate this on substantially longer navigation routes.

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

Related papers

Discussion (0)

Sign in to comment.