NaviWM is a framework designed to enhance robot navigation in human-centric environments by integrating Large Language Models (LLMs) with formal logic. While LLMs excel at high-level reasoning, they often lack physical grounding and logical consistency, leading to "hallucinations" or unsafe spatial decisions. NaviWM addresses these risks in "social navigation"—where robots must respect human comfort and safety—by grounding LLM reasoning in a structured spatial-temporal world model and a verifiable deductive "Chain-of-Thought" (CoT) process.
Methods
The NaviWM architecture is built upon three core pillars:
1. Spatial-Temporal World Model
NaviWM transforms raw sensor data into a structured graph representation rather than feeding unstructured text to the LLM.
- Vertices: Represent agents (robot and humans) with attributes such as position, velocity, and activity (e.g., "walking," "watching phone").
- Edges: Capture spatial relationships (distances) and temporal dynamics (changes over time).
2. Deductive Chain-of-Thought (CoT)
Unlike standard heuristic CoT, NaviWM employs Gentzen-style natural deduction. This forces the LLM to follow a multi-step logical inference process where each step must satisfy First-Order Logic (FOL) predicates, ensuring that navigation decisions are mathematically verified against safety and social norms.
3. Hierarchical Constraint Satisfaction
The system manages social complexity through four levels of compliance. Level 1 represents the ideal state, satisfying activity-aware distancing and general proxemics. In dense crowds where ideal distancing is impossible, the system gracefully degrades to Level 4, which prioritizes basic collision avoidance and goal reaching to prevent the "frozen robot" problem.
Evaluation and Results
The framework was evaluated using an OpenAI Gym-based social navigation simulator across "moderate" (5 humans) and "complex/dense" (10 humans) scenarios. NaviWM was compared against state-of-the-art models including LLaMA-3, DeepSeek-V3, and GPT-4.
Key results include:
- Success Rate (SR): In dense 10-human environments, NaviWM achieved a 0.7 success rate, whereas vanilla LLMs like GPT-4 and LLaMA-8B recorded a 0.0 SR, failing to navigate successfully.
- Social Compliance: NaviWM reduced "Uncomfortable Frequency" (UF) by approximately 50% compared to GPT-4o in dense crowds.
- Ablation Success: Integrating the World Model with DeepSeek-V3 improved its success rate from 0.4 to 0.5 and significantly enhanced social compliance metrics, proving that structured representation is as vital as the reasoning engine itself.
Key Insights
- Logic as a Safety Rail: By requiring the LLM to verify trajectories against a logical proof rather than just calculating them, the system's reasoning becomes interpretable and verifiable.
- Structured Representation: The world model provides the "physical common sense" that LLMs inherently lack, which is crucial for physical grounding.
- Robustness through Degradation: Hierarchical logic allows the robot to maintain safety even when social norms cannot be fully met, ensuring operational continuity in chaotic real-world settings.
Limitations and Future Work
Current limitations include a dependency on accurate perception data and the latency introduced by multi-step LLM queries. Future work may involve integrating Vision-Language Models (VLMs) to better capture human intent and optimizing deduction speed for real-time responsiveness.