ESCAPE: Episodic Spatial Memory and Adaptive Execution Policy for Long-Horizon Mobile Manipulation

ESCAPE addresses a critical bottleneck in robotics: the 'forgetting' and rigid execution issues that occur when a robot must perform complex, multi-step tasks across large indoor spaces. With 5 forks in just 2 days despite 0 stars, it indicates immediate peer interest in the robotics research community. However, its defensibility is limited to its specific architectural approach (episodic spatial memory); it lacks the data moat or hardware lock-in required for a higher score. It competes in a 'red ocean' of embodied AI research where frontier labs like Google DeepMind (RT-2, SayCan), Meta (Habitat-Ego), and NVIDIA (Isaac Lab) are aggressively building foundation models for robotics. The primary risk is that frontier labs will solve long-horizon planning through massive-scale 'World Models' or VLMs that inherently handle spatial reasoning without needing the specific episodic memory architecture proposed here. As a reference implementation for a paper, its primary value is as a benchmark or a module for other researchers rather than a production-ready system. Displacement is likely within 18-24 months as unified VLA (Vision-Language-Action) models improve at zero-shot long-horizon reasoning.