Vector-Quantized Geodesics Enable Spatial Reasoning in LLM Agents

This architecture replaces direct LLM action prediction with vector-quantized geodesic trajectories used as tools. An LLM orchestrator selects these pre-defined spatial tools based on natural language descriptions, separating low-level trajectory execution from high-level reasoning.