RoboTALES: Learning Reasoning-Guided Robot Policies via Task-Aligned Simulated Futures

RoboTALES is a single-stage framework for extracting robot policies from imagined futures generated by pretrained video models. A hierarchical LLM planner decomposes complex tasks into subgoals that guide future generation, while a VLM critic evaluates imagined rollouts and supplies reward-based feedback. The authors report more temporally consistent, task-aligned rollouts and coherent actions on manipulation tasks from RoboCasa and LIBERO10, with especially strong gains on long-horizon tasks. Code and models are publicly released, although the supplied abstract does not provide numerical results, baselines, or task-level success rates.