Researchers present MaP-WAM, a framework for non-Markovian robotic manipulation that decomposes memory-dependent world-action modeling into memory-grounded planning and plan-conditioned execution. The system represents memory as episodic records combining language and visual context, then converts these records into compact plans that guide action execution through a World-Action-Progress model predicting action chunks with adaptive segment transitions. The approach reaches an 83.3% success rate on RMBench benchmarks and 78.0% success on real-robot tasks, while keeping executor latency constant even as task history grows longer.
