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

TL;DR

StressDream steers video world models by optimizing initial noise to enhance robust policy evaluation.

cs.CV 🔴 Advanced 2026-05-30 48 views
Junwon Seo Sushant Veer Ran Tian Wenhao Ding Apoorva Sharma Karen Leung Edward Schmerling Marco Pavone Andrea Bajcsy
video world models policy evaluation noise optimization autonomous driving robotic manipulation

Key Findings

Methodology

StressDream optimizes the initial noise of diffusion-based models to steer generation towards high-impact yet plausible outcomes. It uses a Vision-Language Model for semantic gradients and maintains noise plausibility.

Key Results

  • In experiments on autonomous driving and robotic manipulation, StressDream increased recall for detecting high-impact outcomes from 54% to 94%.
  • StressDream outperformed random sampling methods in task failure detection, significantly improving recall.
  • By steering towards plausible high-impact outcomes, StressDream achieved robust policy evaluation.

Significance

StressDream introduces guided generation of high-impact outcomes in policy evaluation, crucial for fields like autonomous driving and robotic manipulation. It addresses the lack of exploration for critical outcomes in traditional methods.

Technical Contribution

StressDream presents a novel inference-time steering method, optimizing high-dimensional noise with semantic and plausibility objectives, significantly enhancing video world model generation capabilities.

Novelty

This is the first method to steer video world models towards high-impact outcomes by optimizing initial noise, offering a more efficient policy evaluation approach compared to existing methods.

Limitations

  • Optimizing in high-dimensional noise spaces remains challenging, potentially leading to high computational costs.
  • Relies on the accuracy of Vision-Language Models, which may be limited by their understanding capabilities.

Future Work

Future work could explore more efficient noise optimization algorithms and applicability in broader scenarios.

AI Executive Summary

In fields like autonomous driving and robotic manipulation, policy evaluation and improvement rely on accurately imagining possible future outcomes. However, traditional methods often overlook high-impact results. StressDream steers video world models by optimizing initial noise to generate high-impact yet plausible outcomes, enhancing robust policy evaluation.

StressDream uses a Vision-Language Model for semantic gradients and maintains noise plausibility by keeping it within distribution. In experiments, StressDream significantly increased recall for detecting high-impact outcomes in autonomous driving and robotic manipulation, demonstrating its effectiveness in policy evaluation.

Nevertheless, optimizing in high-dimensional noise spaces remains challenging. Future work could explore more efficient algorithms and applicability in broader scenarios.

Deep Analysis

Background

Video world models show promise in policy evaluation and improvement, simulating future observations without costly real-world interactions. However, traditional methods rely on nominal imaginations, potentially missing high-impact outcomes.

Core Problem

Traditional video world models often overlook high-impact outcomes in policy evaluation, leading to less robust policy improvement. How to steer models to generate plausible high-impact outcomes at inference time is a key challenge.

Innovation

StressDream steers generation towards high-impact outcomes by optimizing initial noise in diffusion models. It uses a Vision-Language Model for semantic gradients and maintains plausibility by keeping noise within distribution.

Methodology

  • �� Use diffusion models as video world models
  • �� Optimize initial noise to steer generation
  • �� Use Vision-Language Model for semantic gradients
  • �� Maintain plausibility by keeping noise within distribution

Experiments

Experiments conducted in autonomous driving and robotic manipulation using Vista and Ctrl-World models, optimizing noise to generate high-impact outcomes. Evaluations include recall rates and video quality.

Results

StressDream significantly increased recall for detecting high-impact outcomes in autonomous driving and robotic manipulation, from 54% to 94%. Results show StressDream can generate plausible high-impact outcomes.

Applications

StressDream has direct applications in autonomous driving and robotic manipulation, enhancing robust policy evaluation and reducing potential failures.

Limitations & Outlook

While StressDream performs well in policy evaluation, optimizing in high-dimensional noise spaces remains challenging, potentially leading to high computational costs.

Plain Language Accessible to non-experts

Imagine you're cooking in a kitchen. You have a recipe, but it doesn't tell you all the possible mistakes. StressDream is like a smart assistant that predicts mistakes you might make, like adding too much salt or not cooking long enough. It adjusts details to ensure your dish is both delicious and safe.

ELI14 Explained like you're 14

Imagine you're playing a game where you need to predict the enemy's moves. StressDream is like a super cheat code that helps you predict the enemy's next move, so you can prepare in advance. It adjusts some settings to give you an advantage in the game.

Glossary

Video World Model

A generative model that simulates future observations, often used for policy evaluation and improvement.

Used to generate future scenarios for policy assessment.

Diffusion Model

A generative model that produces samples by gradually denoising.

Serves as the foundation for video world models.

Vision-Language Model

A model that combines visual and language information for reasoning.

Provides semantic gradients to guide generation.

Nominal Imagination

The standard generation process that does not consider high-impact outcomes.

Commonly used in traditional policy evaluation.

High-impact Outcome

Outcomes critical for policy evaluation and improvement.

The target that StressDream aims to steer generation towards.

Open Questions Unanswered questions from this research

  • 1 How to optimize in even higher-dimensional noise spaces? Current methods have computational efficiency limitations.
  • 2 How to improve Vision-Language Model accuracy to better guide generation?

Applications

Immediate Applications

Autonomous Driving

In autonomous driving, StressDream can help predict potential collisions or other hazards, enhancing vehicle safety.

Long-term Vision

Intelligent Robotics

In the future, StressDream could be used in intelligent robots to better predict and avoid failures in complex environments.

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/.

cs.CV cs.AI cs.LG cs.RO