Chameleon: Episodic Memory Architecture for Long-Horizon Robotic Manipulation Under Occlusion

arXiv·low signalXBlueskyLinkedInCopy link

Chameleon introduces an episodic memory system for robotic manipulation that addresses perceptual aliasing — when occlusion and state changes make the same observation arise from different underlying states, breaking the Markov assumption. The memory architecture allows robots to recall and reason about past observations to disambiguate current situations during long-horizon tasks. While focused on robotics, the episodic memory design patterns are transferable to any agent operating in partially observable environments.