Memory as Plans: World-Action Modeling with Memory-Grounded Planning
Non-Markovian manipulation tasks typically force robotic policies to trade history coverage against real-time latency. MaP-WAM resolves this tension by decoupling world-action modeling into memory-grounded planning and plan-conditioned execution. Rather than feeding cumulative multimodal histories to the controller, it converts episodic records into compact segment plans, while a World-Action-Progress model executes them within a fixed context window. The framework posts an 83.3% success rate on RMBench and 78.0% on real-robot tasks with near-constant executor latency.
Why it's worth reading
It tackles the latency penalty inherent in long-horizon robotic memory, offering a decoupled architecture that retains fine-grained episodic context without bloating the executor's real-time inference cost.