From VPP to VPP2: Your WAM May Not Generalize as Much as You Think

Today's best video generation models are trained on videos from across the entire internet. They have seen kitchens, construction sites, sports fields, operating tables; they have seen human hands, robotic arms, cats, and trucks. These models carry something very valuable: generalization ability.

After interfacing with actions, we similarly hope that robots will show significantly improved generalization when facing unseen tasks.

But Yanjiang, an author of the VPP series, mentioned a counterintuitive phenomenon when communicating with us: WAM has not actually achieved the ideal generalization effect on unfamiliar scenes and unfamiliar tasks. (It is worth noting that Video prediction policy is a work from late 2024, earlier than the current WAM wave. But by 2026, WAM has not actually been optimized to the ideal effect.)

Look through the project pages of previous WAM papers. You will notice a strange pattern: the video predictions they show almost invariably fall within robotic arm scenes on which they were post-trained. A few familiar objects are placed on the table, and the robotic arm reaches over in a familiar posture. In the few examples showing human hands, the hands only make slight movements in the frame.

In the article, VPP2 conducts a comparative experiment: on standard LIBERO, some mainstream WAM models achieve very high success rates, but when switched to LIBERO-OOD, which tests generalization ability, the scores drop sharply. This gap shows that high success rates on familiar tasks do not prove that video priors have been brought into unfamiliar tasks.