Video generative models are increasingly seen as a promising basis for robotic planning, as they can predict future visual rollouts from observations and task instructions. This capability could let robots imagine the consequences of their actions before executing them.
However, the abstract points out a key shortcoming: the training objectives for these models typically lack explicit constraints on robot actions. In other words, the generated videos may look plausible but not correspond to actions a robot can actually perform.
The proposed EVA method addresses this by using inverse dynamics rewards to align the video world model with executable robot actions. This means the model is trained not just to predict pixels, but to produce rollouts that are consistent with feasible motor commands.