CLAW: Continuous Latent Action World Models via Adversarial Regularization

This research introduces CLAW, a fully end-to-end self-supervised framework for learning world models and continuous latent action representations directly from action-free videos. The method uses adversarial latent regularization to ensure the latent space captures action structure rather than future observations, while a diffusion-based component models environment dynamics as a distribution over future states. Trained solely on observation sequences without action annotations, CLAW supports behavior cloning for imitation learning and generates executable action sequences for goal-directed planning. Experiments across diverse tasks and embodiments demonstrate that CLAW produces semantically meaningful latent actions, facilitates effective action transfer, and outperforms existing methods in planning and imitation from visual observations alone.