MaP-WAM Robot Framework Reports 83.3% on RMBench, 78% on Real Bot

A new arXiv preprint opens with a familiar complaint about robot policies: "Mainstream robotic policies often adopt a Markovian formulation, but many complex real-world manipulation tasks are inherently non-Markovian, requiring long-horizon memory beyond the current observation." The fixes so far, per the authors, are language summaries or growing visual windows, both trading history coverage against execution speed.

MaP-WAM, from Sizhe Zhao and co-authors, treats memory as a plan instead. The system "represents memory as completed segment records containing language instructions and sparse visual context, and converts this episodic memory into compact plans." A separate World-Action-Progress model then executes each plan, jointly predicting action chunks and how far along it is, and re-aligning when the observation drifts from the plan.

The headline numbers: 83.3% success on RMBench and 78.0% on real-robot tasks. The executor's context length stays fixed as task history piles up, and the paper reports "approximately constant executor inference latency as task history grows."

The abstract does not name the baselines the 83.3% is beating, describe the real-robot suite behind the 78.0%, or say what the planning stage itself costs as episodic memory accumulates.