R2-WAM: data-free continual learning for world action models via rectified generative replay

Loading video

Loading video

R2-WAM lets a world action model rehearse its own regenerated rollouts of past tasks, truncated at task completion by an agentic evaluator and filtered for action-observation consistency. Learning 10 tasks in sequence lifts earlier-task success from 28.5% to 89.5%; on LIBERO and five real-robot tasks it cuts forgetting over 90% versus sequential fine-tuning, with no stored old data.