MaP-WAM stores robot memory as completed segment records and turns them into plans, rather than replaying full history

arXiv (2609.11561)·low signalXBlueskyLinkedInCopy link

Most robotic policies are Markovian while real manipulation tasks are not, and the usual fixes — language summaries or growing visual windows — either lose fine-grained visual evidence or force a trade between history coverage and execution speed. MaP-WAM splits the problem: episodic memory is kept as completed segment records holding a language instruction plus sparse visual context, compressed into a plan for the next segment, and a World-Action-Progress model executes that plan over an unknown duration. The executor never conditions on the full history, which is the same context-budget argument agent framework authors keep rediscovering outside robotics.