A new arXiv preprint introduces DreamCan, a hierarchical policy designed to make video models useful for real robot control. The authors note that while video models provide strong priors for generalizable manipulation, the visual futures they generate are not always executable by a physical robot. DreamCan addresses this by grounding video models through sparse visual planning, selecting key waypoints rather than dense action sequences.
The paper argues that this sparse planning approach helps reconcile what looks plausible in a video with what can actually be achieved. By focusing on critical intermediate states, the policy can guide long-horizon tasks without relying on the video model to produce every low-level action. The abstract does not include experimental details, so the reported mechanism is limited to the described architecture and motivation.
As a preprint, the work has not yet undergone peer review, and the abstract leaves open questions about benchmark performance and comparisons to other methods. Still, the core idea—using sparse visual plans to anchor video-based predictions—offers a practical direction for closing the gap between visual imagination and robotic execution.