World Action Models (WAMs) have become a common way to adapt video foundation models to robotic control: they learn to predict both video and actions. But according to a new arXiv preprint, task-level instructions alone often do not specify exactly what the robot should do next. The paper, Atomic-WAM, argues that this ambiguity is a core problem for control.
The proposed approach is execution-grounded atomic-motion reasoning. Instead of relying only on high-level task descriptions, the model reasons about the next physical interaction at a finer granularity, grounded in execution. The abstract is truncated in the source, so full evaluation details are not available in this digest.