Comparison of Vision Language Action using goal state as action tokens

Vision–Language–Action (VLA) models differ in how they represent actions. Traditional VLA approaches predict low-level control commands or trajectories directly. In contrast, goal-state-as-action-token methods encode actions as desired future states (e.g., object poses or scene configurations). This shifts the problem from how to act to what outcome to achieve, improving abstraction, compositionality, and generalization across tasks while relying on a downstream controller to execute the goal.