Seonglae ChoVLA is closer to a reactive policy, whereas WAM combines a world model with an action model.
WAM aims to directly learn the joint distribution over a future state and an action , given the current observation and a language instruction . Here, denotes the agent’s inputs (e.g., vision and proprioception), and is the next environment state resulting from the action. While a VLA model typically optimizes only the negative log-likelihood of the action, , WAM optimizes a state-inclusive objective, . This encourages the model to internalize physical causal structure beyond mere control, and to learn useful physical commonsense even from internet-scale video data where actions are not explicitly labeled.
