StressDream: Steering Video World Models for Robust Policy Evaluation and Improvement

2026-05-29

AI Summary

Existing video world models rely on conventional imagined trajectories for policy evaluation, struggling to efficiently generate high-impact yet plausible future scenarios—such as task failures—which leads to insufficient identification of risky actions. To address this limitation, this work proposes a dual-objective optimization mechanism at inference time that precisely steers the initial noise in diffusion-based video world models by combining semantic objectives—guided by gradients from a vision-language model—with constraints on the noise distribution. This approach enables text-specified generation of future scenarios that are both semantically coherent and within the data distribution. It represents the first method capable of controllably guiding high-dimensional imagined trajectories, successfully identifying high-risk actions in autonomous driving and robotic manipulation tasks, thereby significantly enhancing policy robustness.

Abstract

Video world models (WMs) have shown promise for policy evaluation and improvement by imagining realistic future observations conditioned on ego-robot actions. While WMs can model distributions over futures, policy evaluation and improvement typically rely on nominal imaginations, which can miss high-impact outcomes of robot actions unless prohibitively many samples are drawn. To enable robust policy evaluation and improvement over WM imaginations, we propose StressDream, which steers imaginations toward high-impact yet plausible outcomes specified at inference time by optimizing the initial noise of diffusion-based WMs. However, optimizing high-dimensional noise is challenging: the optimization must reason about nuanced, scene-dependent target events in generated videos while avoiding out-of-distribution (OOD) noise that yields implausible imaginations. We address this with two complementary objectives: a semantic objective with a Vision-Language Model that provides informative gradients by reasoning about the generated video, and a plausibility objective that prevents the optimized noise from drifting OOD. With state-of-the-art video world models for autonomous driving and robotic manipulation, we show that StressDream effectively steers imaginations toward high-impact yet plausible outcomes specified by text at inference time, such as task failures, enabling robust policy evaluation and improvement by identifying actions whose plausible futures include undesirable outcomes. Video results are available at https://junwon.me/StressDream/.

Problem

• video world models
• policy evaluation
• high-impact outcomes
• robustness
• plausible imagination

Innovation

• video world models
• diffusion models
• robust policy evaluation
• noise optimization
• vision-language model