MemoGuard Enhances Robot Navigation Safety in Communication-Limited Environments

MemoGuard is a novel adaptive runtime system designed to enhance the safety and efficiency of robots operating in communication-limited, mission-critical environments, such as disaster inspection or search-and-rescue. In these scenarios, robots often rely on episodic memory reuse as a low-cost fallback mechanism. However, high retrieval similarity does not guarantee the validity or safety of a retrieved action, leading to "memory traps" where an action might be unsafe due to changed conditions like topology, battery levels, or unreliable prior outcomes. MemoGuard addresses this by introducing a lightweight validation process. Before reusing an episodic memory, it checks against predefined contracts related to topology, resources (e.g., battery margin), and outcome reliability. Only if validation fails does the system invoke more computationally expensive local reasoning as a fallback. This approach strikes a balance between safety and efficiency. In a simulated corridor-inspection environment, MemoGuard demonstrated significant improvements: it reduced battery safety violations by 76.6% compared to similarity-only reuse and decreased fallback calls by 21.4% compared to always invoking reasoning. On an NVIDIA Jetson AGX Xavier, this translated to substantial savings in computational overhead (3.67 seconds and 36.97 Joules per trial). The system is open-source, providing a practical solution for robust robot navigation.