Nav-CoT-110K: Synthetic CoT Dataset

• Nav-CoT-110K is a synthetic dataset that provides 110K chain-of-thought annotated navigation trajectories for structured reasoning in embodied tasks.
• It integrates low-level visual and tactile inputs with high-level semantic reasoning to support both supervised and reinforcement learning approaches.
• It underpins modular reasoning pipelines and dual-layer control systems, significantly improving navigation success and trajectory efficiency.

Nav-CoT-110K is a large-scale synthetic dataset of 110,000 chain-of-thought (CoT) annotated trajectories designed to support structured reasoning in embodied navigation and complex multimodal tasks. Serving as a foundation for modular reasoning pipelines and enabling stepwise decision-making in both simulation and real-world environments, Nav-CoT-110K underpins recent advances in embodied foundation models such as Nav-R1. The dataset aligns low-level action spaces with high-level semantic reasoning, providing explicit chains of visual and linguistic cues linked to navigation commands and forming a standardized template for cold-start policy initialization and reinforcement learning refinement.

1. Dataset Structure and Annotations

Nav-CoT-110K consists of 110K synthetic examples, each pairing natural language navigation instructions with egocentric inputs—specifically RGB-D images and point cloud data. For every input, the CoT annotation unfolds a step-by-step reasoning process that culminates in selecting a discrete action for navigation. The output of each sample uses a structured markup format:

<think> …reasoning steps… </think><action> …chosen action… </action>

This strict separation ensures clarity and machine-readability, facilitating disentanglement of high-level reasoning from low-level control policies. The dataset encodes joint linguistic-visual alignment: reasoning traces explicitly reference semantic cues from observations and link these to navigation intentions, enabling models to learn coherent mappings between what is perceived, how it is interpreted, and the actions that follow. The annotation paradigm is architected to support both supervised learning and reinforcement learning objectives.

2. Role in Embodied Reasoning Models

Nav-CoT-110K is central to cold-start initialization of embodied foundation models, notably Nav-R1 (Liu et al., 13 Sep 2025). During initial fine-tuning, the model is trained to generate outputs strictly conforming to the markup template. This bootstrapping phase equips the model with explicit multimodal reasoning capabilities. Subsequently, reinforcement learning stages refine the policy with task-specific objectives, leveraging the data's structured chains for format adherence, semantic consistency, and trajectory fidelity. The dataset's design is architected for scalability, supporting larger backbone models and enabling transfer across varied embodied scenes.

3. Selective Chain-of-Thought Reasoning

Empirical meta-analysis demonstrates that chain-of-thought reasoning yields pronounced performance gains primarily in tasks involving explicit symbolic or mathematical operations, as observed in (Sprague et al., 18 Sep 2024). For tasks lacking intermediate formal operations (often indicated by the absence of symbols such as ' = '), direct answer generation performs comparably; CoT's advantage is concentrated in settings requiring stepwise decomposition (e.g., spatial reasoning, logic, or equation solving). Nav-CoT-110K, by structuring reasoning chains, aligns with this selective paradigm—triggering detailed stepwise reasoning only when the input or required reasoning contains explicit symbolic manipulations. This selectivity optimizes inference cost and reduces risk of chain-errors in domains lacking intermediate computational complexity.

4. Integration with Neuro-Symbolic and Tool-Augmented Frameworks

Frameworks adopting Nav-CoT-110K increasingly embrace hybrid neuro-symbolic paradigms. For symbolic and mathematical planning, the separation of planning and execution stages enables models to generate formal plans and delegate execution to dedicated solvers, surpassing raw chain-of-thought execution. In embodied navigation, structured chain-of-thought traces can interface with external modules—such as spatial planners or 3D reconstruction experts (as in OCC-MLLM-CoT-Alpha (Wang et al., 7 Apr 2025))—further boosting task performance under occlusion and ambiguity. This hybridization reflects a broader trend towards modular composition: high-level chain-of-thoughts are used for interpretive planning, while specialized tool or solver modules handle computationally demanding execution steps.

5. Training and Reinforcement Learning Methodology

Nav-CoT-110K underpins reinforcement learning approaches, most notably Group Relative Policy Optimization (GRPO) (Liu et al., 13 Sep 2025). Post-initialization, the policy is refined through sampling multiple candidate outputs per observation and scoring them according to:

• Format Reward ($R_{Format}$): Enforces strict conformity to the template structure.
• Understanding Reward ($R_{understanding}$): Aggregates exact answer matching and semantic alignment using metrics such as CLIPScore.
• Navigation Reward ($R_{navigation}$): Measures path and endpoint accuracy, using $D_F(\cdot)$ for path fidelity and Euclidean distance to the true endpoint.

Mathematically, the policy update is governed by:

$$J_{GRPO}(\theta) = \mathbb{E}_c \left[ \frac{1}{G} \sum_{i=1}^G \min \left\{ \frac{\pi_\theta(o^i|q)}{\pi_{\theta_{\text{old}}}(o^i|q)} \cdot \hat{A}_i, \operatorname{clip}\left(\frac{\pi_\theta(o^i|q)}{\pi_{\theta_{\text{old}}}(o^i|q)}, 1-\epsilon, 1+\epsilon\right) \cdot \hat{A}_i \right\} - \beta D_{\text{KL}}(\pi_\theta || \pi_{\text{ref}}) \right]$$

where $\hat{A}_i$ denotes normalized advantage, $\beta$ is the KL penalty, and $\epsilon$ controls policy update bounds. This reinforcement protocol leverages Nav-CoT-110K’s structured annotations for multi-dimensional reward calculation.

6. Dual-System Reasoning and Control Paradigms

Nav-R1 and related models trained with Nav-CoT-110K implement a dual-system architecture (termed "Fast-in-Slow" (Liu et al., 13 Sep 2025)):

• Slow System (System 2): Aggregates long-horizon semantic history and language instructions to generate latent guidance every ~3 steps.
• Fast System (System 1): Processes high-frequency sensory input and latent guidance to output reactive control sequences for immediate execution.

This decoupling enables robust semantic planning without sacrificing real-time performance. Asynchronous coordination ensures global coherence (via slow, deliberative reasoning) while maintaining responsive control in dynamic environments.

7. Empirical Evaluations and Impact

Nav-R1 trained with Nav-CoT-110K demonstrates over 8% average improvement in navigation success and trajectory efficiency on benchmarks such as VLN-CE, RxR-CE, and HM3D-OVON (Liu et al., 13 Sep 2025). Real-world deployments achieve navigation error (NE) as low as ~1.2 and normalized success rates exceeding 1.0 across indoor settings, confirming robustness under computation-constrained scenarios. In related vision-language tasks, integration with multi-modal chain-of-thought annotation (as in OCC-MLLM-CoT-Alpha (Wang et al., 7 Apr 2025)) yields 15–17% decision score boosts for occlusion recognition.

8. Research Implications and Directions

Nav-CoT-110K substantiates the selective utility of chain-of-thought prompting—most poignant for symbolic, spatial, or stepwise reasoning domains. Methodological innovations stemming from this dataset include structured annotation paradigms, multi-dimensional reinforcement learning protocols, and dual-layer reasoning/control. The dataset’s modularity supports neuro-symbolic and tool-augmented models, and its selective triggering of CoT chains offers efficiency gains and error reduction. Future directions explore richer intermediate computation paradigms, integrating recursive verification, agent-based compositional reasoning, and selective invocation tied to task characteristics, as outlined in (Sprague et al., 18 Sep 2024).