Researchers have published DSWAM, a robot control model that switches between quick physical execution and deliberate language-level planning depending on how complex the task at hand is.
Existing approaches split into two camps. Vision-Language-Action (VLA) models are good at parsing instructions but lean heavily on language reasoning for every step. World Action Models (WAMs) skip that and use video-based world modeling to ground actions in physics — but they struggle to break a vague command like "clean the kitchen" into executable steps. DSWAM tries to get both: a System 1 WAM executor runs by default, and a System 2 vision-language planner kicks in only when a task needs decomposition. The planner reads visual history and a global prompt, outputs subtasks, and hands them off to the executor one at a time. To make this usable on real hardware, the team added TensorRT acceleration, asynchronous execution, and real-time chunking so the policy does not stall the robot mid-motion.
The dual-system framing borrows from cognitive science — fast and slow thinking applied to robot arms — and it sidesteps a genuine problem: most WAM-vs-VLA comparisons to date have been apples-to-oranges, differing in training data, robot hardware, and task criteria. DSWAM is evaluated under the DeMaVLA benchmark, which holds those variables fixed, giving the field a cleaner number to argue over. That controlled comparison is arguably the paper's most useful contribution, independent of whether DSWAM itself wins.
Robot manipulation research is crowded with architectures that benchmark well and deploy poorly; the engineering work here — latency reduction, chunked inference — suggests the authors are at least aware of that gap, even if closing it at household scale remains a different problem entirely.