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

RoboTALES, from the University of California, San Diego, reframes robot control as task-aligned imagination by coupling a pretrained Stable Video Diffusion backbone with hierarchical reasoning and policy learning. A Gemini 2.5 Pro LLM planner decomposes each instruction into 2 to 5 ordered subgoals, the resulting plan conditions the video generator, and a frozen VLM critic provides DDPO-style reward feedback that steers the denoising dynamics toward semantically faithful futures rather than merely realistic ones. Those learned decoder features then condition a 1D action diffusion UNet, with action gradients flowing back into the video generator in a single-stage joint objective. On RoboCasa, the method reaches 64% average success across 24 manipulation tasks, improving over VideoPolicy and the strongest prior baseline, and on LIBERO10 it achieves 97% average success across 10 tasks, versus 94% for VideoPolicy. The paper reports especially strong gains on structured long-horizon behaviors such as turning, pressing, and pick-and-place, while ablations show that planner injection alone is insufficient without joint training and critic feedback. Additional analysis finds the critic reward is calibrated to success with AUROC 0.72, and joint training also improves general video generation fidelity on UCF101, reducing Fréchet distance from 0.6715 to 0.2437 and MMD-RBF from 0.2419 to 0.0631.