PaperlayerView exact revision on arXiv

Source-linked AI summary

Nav-R1: Reasoning and Navigation in Embodied Scenes

Qingxiang Liu, Ting Huang, Zeyu Zhang, Hao Tang

arXiv:2509.10884v1cs.ROcs.CV

TL;DR

Embodied navigation must reconcile unstable reasoning with the need for long-horizon semantics and low-latency control. Nav-R1 combines structured CoT initialization, GRPO rewards, and Fast-in-Slow reasoning, and reports consistent benchmark improvements alongside mobile-robot deployment under limited onboard resources.

  • Problem

    Existing embodied-navigation approaches can produce incoherent reasoning traces and struggle to balance long-horizon semantic reasoning with low-latency reactive control.

  • Method

    Nav-R1 uses Nav-CoT-110K for cold-start initialization, GRPO with format, understanding, and navigation rewards, and a Fast-in-Slow dual-system reasoning scheme.

  • Results

    Nav-R1 consistently outperforms prior methods across instruction-following and object-goal navigation benchmarks, with higher success rates and trajectory efficiency and fewer navigation errors.

  • Takeaways & Limitations

    Nav-R1 reports consistent improvements in navigation success, trajectory fidelity, and reasoning coherence across VLN, ObjectNav, dialogue, planning, and reasoning benchmarks.

  • Takeaways & Limitations

    Nav-R1 relies on RGB-D and language inputs, while current deployment depends on cloud inference and the dataset does not fully capture real-world complexity.

Abstract

from arXiv · show

Embodied navigation requires agents to integrate perception, reasoning, and action for robust interaction in complex 3D environments. Existing approaches often suffer from incoherent and unstable reasoning traces that hinder generalization across diverse environments, and difficulty balancing long-horizon semantic reasoning with low-latency control for real-time navigation. To address these challenges, we propose Nav-R1, an embodied foundation model that unifies reasoning in embodied environments. We first construct Nav-CoT-110K, a large-scale dataset of step-by-step Chains-of-Thought (CoT) for embodied tasks, which enables cold-start initialization with structured reasoning. Building on this foundation, we design a GRPO-based reinforcement learning framework with three complementary rewards: format, understanding, and navigation, to improve structural adherence, semantic grounding, and path fidelity. Furthermore, we introduce a Fast-in-Slow reasoning paradigm, decoupling deliberate semantic reasoning from low-latency reactive control for efficient yet coherent navigation. Extensive evaluations on embodied AI benchmarks demonstrate that Nav-R1 consistently outperforms strong baselines, with over 8% average improvement in reasoning and navigation performance. Real-world deployment on a mobile robot further validates its robustness under limited onboard resources. Code: https://github.com/AIGeeksGroup/Nav-R1. Website: https://aigeeksgroup.github.io/Nav-R1.

I. INTRODUCTION

Nav-R1 addresses incoherent reasoning and the difficulty of combining long-horizon semantics with low-latency control in embodied navigation. It combines structured CoT initialization, reward-based reinforcement learning, and Fast-in-Slow reasoning, with reported gains on benchmarks and real-robot deployment.

  • I. INTRODUCTION: Nav-R1 targets unstable reasoning traces and the unresolved need to balance long-horizon semantic reasoning with low-latency reactive control.
  • I. INTRODUCTION: Nav-R1 combines Nav-CoT-110K cold-start data, GRPO rewards for format, understanding, and navigation, and a Fast-in-Slow dual-system design.The dataset supplies step-by-step reasoning trajectories, while the rewards target structural adherence, semantic grounding, and path fidelity.
  • I. INTRODUCTION: Nav-R1 integrates reasoning, planning, dialogue, and navigation within a unified embodied foundation model.
  • I. INTRODUCTION: Nav-R1 reports consistent improvements over prior methods in navigation success, trajectory fidelity, and reasoning accuracy across embodied benchmarks.The supplied passages state that the model outperforms prior state-of-the-art methods and reports an average improvement of 8% across several task types.
  • I. INTRODUCTION: The framework is also evaluated on a physical mobile robot under limited onboard computation using cloud-based inference.

II. RELATED WORK

Related work spans embodied understanding, multimodal perception-language-action models, and end-to-end navigation. Nav-R1 builds on these directions while combining structured reasoning with dual-speed control across diverse embodied datasets.

  • II. RELATED WORK: Embodied understanding integrates multimodal sensory data and linguistic instructions so agents can perceive, reason, and act in 3D environments.
  • II. RELATED WORK: Recent embodied agents use large language models to connect perception, grounding, and action planning, with some pursuing unified competence without task-specific submodules.
  • II. RELATED WORK: Embodied navigation includes ObjectNav and VLN, requiring multimodal instructions to be translated into continuous motion through unstructured 3D scenes.
  • II. RELATED WORK: Recent navigation methods explore end-to-end vision-language-action models, reinforcement fine-tuning, chain-of-thought training, and active exploration.
  • II. RELATED WORK: Nav-R1 draws on public 3D vision-language datasets covering varied scenes, task types, instruction modalities, and environment settings.

B. Synthetic Dataset

Nav-CoT-110K is constructed from multimodal navigation examples that pair egocentric observations, instructions, feasible actions, and structured reasoning. A filtering pipeline retains 110K trajectories aligned with feasible navigation paths.

  • Synthetic Dataset: 110K step-by-step Chain-of-Thought trajectories form Nav-CoT-110K, providing structured reasoning aligned with multimodal observations for cold-start training.The dataset bridges perception, language, and action before reinforcement learning.
  • Synthetic Dataset: Instructions come from VLN and ObjectNav benchmarks, including R2R, R2R-CE, RxR-CE, SOON, and OVON.The source tasks provide diverse navigation instructions grounded in embodied-AI benchmarks.
  • CoT Data Engine: The CoT data engine combines navigation instructions, egocentric visual inputs, feasible action options, and explicit output formatting to prompt structured reasoning.Gemini 2.5 Pro generates reasoning and corresponding actions or answers in standardized tags.
  • Synthetic Dataset: Approximately 115K raw CoT examples are filtered by rule-based consistency checks and trajectory verification, leaving 110K refined examples.Verification cross-validates actions against feasible navigation paths.

C. Cold Start Stage

The cold-start and policy-training pipeline initializes Nav-R1 with structured reasoning before reinforcement learning. GRPO then evaluates multiple candidate outputs with complementary format, understanding, and navigation rewards.

  • Cold Start Stage: Nav-CoT-110K supervised fine-tuning initializes structured <think>...</think><action>...</action> outputs grounded in corresponding navigation actions.This cold-start stage is intended to stabilize subsequent reinforcement learning.
  • Policy Samples: Nav-R1 samples multiple candidate answers or actions from multimodal inputs, then evaluates them with task-specific reward functions.Candidates support either scene understanding or navigation decisions.
  • Format Reward: Format reward enforces machine-parseable reasoning-decision templates and separates reasoning from final predictions.Supported templates use either answer or action tags after the reasoning block.
  • Understanding Reward: Understanding reward combines exact answer correctness with semantic alignment to assess genuine 3D-scene comprehension.The design targets both factual errors and semantically irrelevant outputs.
  • Navigation Reward: Navigation reward combines trajectory fidelity with endpoint accuracy to optimize both path alignment and successful goal reaching.The path component compares predicted and reference trajectories, while the endpoint component penalizes final-location deviation.
  • Policy Optimization: GRPO normalizes rewards into relative advantages and updates the policy with a clipped objective regularized by a KL term.The three rewards jointly balance structural validity, semantic understanding, and action execution.

E. Fast-in-Slow Reasoning

Fast-in-Slow couples deliberate semantic reasoning with rapid action execution to balance long-horizon planning and low-latency control. A slow system supplies stable latent guidance while a fast system converts frequent multimodal observations into short-horizon actions.

  • Fast-in-Slow tightly couples deliberate semantic reasoning with rapid action execution for dynamic embodied navigation.The design addresses the tension between accurate long-horizon planning and low-latency control.
  • Slow reasoning: The slow system processes multimodal observations and historical context to produce latent features encoding scene semantics, temporal dependencies, and global navigation goals.
  • Fast reasoning: The fast system fuses RGB, depth, and point-cloud inputs with slow-system latent features to predict short-horizon actions using lightweight reused transformer blocks.The high-frequency policy integrates visual, depth, and 3D geometric cues for real-time control.
  • Asynchronous coordination: A 1:n asynchronous update ratio lets one slow-system update guide n consecutive fast steps, with n ≈3 providing the best balance between semantic fidelity and responsiveness.

A. Benchmarks and Metrics

Nav-R1 is evaluated across navigation, dialogue, reasoning, and planning benchmarks using task-specific standard metrics. Its training pipeline combines synthesized and filtered CoT trajectories, supervised cold-start initialization, GRPO rewards, and parameter-efficient tuning.

  • Benchmarks: The evaluation spans instruction-following, object-goal navigation, open-vocabulary navigation, embodied dialogue, planning, and reasoning tasks.R2R-CE and RxR-CE cover navigation in unseen environments, HM3D covers object-goal navigation, HM3D-OVON tests zero-shot novel categories, and 3D-LLM covers dialogue and planning.
  • Metrics: Navigation is measured with NE, SR, OS, SPL, and nDTW, while language tasks use CIDEr, BLEU-4, METEOR, and ROUGE-L.
  • Data synthesis: Nav-CoT-110K is built from embodied instructions and filtered through rule checks and trajectory verification, leaving 110K high-quality trajectories.
  • Training: Nav-R1 receives supervised fine-tuning on Nav-CoT-110K after initialization from 3D-R1, followed by GRPO using format, understanding, and navigation rewards.
  • Parameter-efficient tuning: LoRA adapters reduce trainable parameters by ∼98% compared with full fine-tuning, with about 142M parameters updated.

C. Main Results

Nav-R1 consistently improves embodied navigation across instruction-following and object-goal benchmarks while preserving dialogue, reasoning, and planning capabilities. Real-world tests show strong performance across varied indoor scenes and hardware-constrained deployment conditions.

  • Embodied dialogue: Nav-R1 maintains dialogue performance close to 3D-R1 while outperforming previous baselines, indicating that navigation-oriented reasoning does not weaken interaction quality.
  • Reasoning and planning: Nav-R1 performs comparably to 3D-R1 on embodied reasoning and on par with prior methods for embodied planning.The planning results indicate coherent multi-step action sequences.
  • Embodied navigation: Nav-R1 consistently outperforms prior methods on instruction-following and object-goal navigation, with higher success and trajectory efficiency and lower navigation error.The comparison covers R2R-CE, RxR-CE, and HM3D-OVON results reported in Tables III and II.
  • Real-world evaluation: Nav-R1 achieves the best real-world results across meeting room, lounge, and corridor environments, with reduced NE and higher SR than prior navigation models.
  • Real-world evaluation: In qualitative real-world evaluations, Nav-R1 reaches diverse targets across clutter, narrow passages, and long-horizon paths in three indoor scenes.The deployment pipeline uses a WHEELTEC R550 with onboard sensing and motor control components.

VI. TEST-TIME EFFICIENCY

Nav-R1 uses cloud-assisted inference because the Jetson Orin Nano has limited onboard resources. Server execution keeps latency near real-time despite the additional overhead of dual-system reasoning.

  • Cloud-assisted inference streams egocentric inputs to a remote server and returns compact navigation commands for execution on the robot.
  • The Jetson Orin Nano’s strict resource limits make large-scale local inference prohibitively slow, motivating the cloud-based deployment design.
  • Nav-R1 achieves ∼95 ms server latency, only slightly slower than NaVid and Uni-NaVid, and the marginal gap does not affect real-time navigation.Server-side execution reduces baseline latency to below 100 ms.
  • The real-world system uses a high-speed WiFi 6E network covering the 200 m2 indoor test area with <10 ms access latency.

APPENDIX

The appendix examines Nav-R1’s dual-system design, reward decomposition, and KL-penalty sensitivity, while documenting limitations in data coverage, modalities, and deployment scalability.

  • Dual-system design: The dual-system design achieves the best trade-off, whereas slow-only reasoning struggles with real-time execution and fast-only control lacks global semantic consistency.The comparison is evaluated on R2R-CE Val-Unseen; asynchronous coordination is identified as effective.
  • Reward decomposition: Removing any of the format, understanding, or navigation rewards degrades performance by respectively harming output structure, semantic grounding, or trajectory fidelity.The reward decomposition is evaluated on HM3D-OVON Val-Unseen.
  • Hyper-parameters: The best KL-penalty trade-off occurs at β = 0.02, while smaller values promote reference-policy divergence and larger values restrict exploration.This sensitivity study uses the RxR-CE Val-Unseen split.
  • Limitations: Nav-CoT-110K is mainly synthesized from existing benchmarks and does not fully capture real-world complexity.This limits the dataset’s coverage of real-world conditions.
  • Limitations: Nav-R1 relies on RGB-D and language inputs, lacks richer modalities such as audio or tactile signals, and currently depends on cloud inference for deployment.The authors identify limited edge-device scalability as a consequence of cloud-based inference.

X. VISUALIZATION

The visualization appendix illustrates Nav-R1’s structured reasoning data, qualitative performance across real-world embodied tasks, and benchmark behavior in navigation environments.

  • CoT visualization: Nav-CoT-110K examples show instructions, egocentric observations, and candidate actions transformed into step-by-step reasoning and navigation decisions.The visualization highlights the role of CoT supervision in stabilizing model training.
  • Real-world navigation: Real-world navigation visualizations combine instructions, RGB observations, depth maps, LiDAR top-down maps, third-person views, and BEV trajectories across three indoor scenes.The scenes include a meeting room, lounge, and corridor.
  • Other embodied tasks: Additional real-world visualizations cover embodied dialogue, reasoning, and planning, highlighting coherent interaction, safe reasoning, and multi-step planning in complex layouts.These results are presented in Fig. 11–13.
  • Simulation benchmarks: Simulation visualizations show long-horizon instruction grounding on VLN-CE R2R and exploration with target-object localization on HM3D ObjectNav.The corresponding results are presented in Fig. 14–16.
Loading 2509.10884v1…