MaP-WAM uses episodic memory to plan robotic manipulation
Compact plans keep the executor context fixed while long-term evidence remains available to the planner.
MaP-WAM targets non-Markov robotic manipulation, where the current observation does not contain everything needed for the next action. It separates memory-grounded planning from plan-conditioned execution rather than repeatedly feeding the executor the full history.
Completed segments retain language instructions and sparse visual context. The planner converts this evidence into a compact next-segment instruction and visual guidance. A World-Action-Progress model predicts actions and progress, using plan-observation alignment to decide segment transitions.
The paper reports 83.3% success on RMBench and 78.0% on its real-robot tasks, with approximately constant executor inference latency as history grows. Fixed executor context does not mean the planner’s full memory is free.
This is a concrete robotics memory interface, not proof of generic human-like learning in all language models. A transfer experiment should compare long-history task success, planner/executor budgets and transition errors on the target embodiment.
Editorial correction (2026-10-13 UTC): Restored the robotics domain, compact-plan mechanism and measured results.