UniW(A)M preprint reports 88.3% success on real-world robot tasks

Researchers report that UniW(A)M achieved 88.3% average success across three real-world manipulation tasks, compared with 76.7% for the strongest baseline. The preprint combines models that predict future scenes with models that generate executable robot actions. Shared model components learn common physical dynamics, while specialized components preserve each interface's capabilities. At deployment, a router learns from a small subset of target-task scenarios to choose an action interface for each execution chunk while keeping the underlying model frozen. The router raised success outside the training distribution from 63.3% to 76.7%.