Robot learns from mistakes to navigate new spaces better

Embodied agents struggle to generalize navigation strategies to unfamiliar environments because they lack structured ways to learn from experience. Robo-Cortex addresses this by automatically converting trajectories—both successful and failed—into natural-language heuristics stored in a Navigation Heuristic Library. The system combines short-term memory for tracking immediate progress with long-term memory that captures reusable principles, then uses a world model and vision-language model to simulate and verify action plans before execution. On three benchmarks (IGNav, AR, AEQA), Robo-Cortex achieved up to 4.16% improvement in success-weighted path length over prior methods, with 15.3% gains when transferring learned heuristics to new environments. Early physical robot experiments confirm the approach works beyond simulation.