DriftWorld: Fast World Modeling through Drifting

2026-07-16

AI Summary

Diffusion models suffer from slow inference due to multi-step sampling, hindering their applicability in large-scale action search for robotic planning. This work proposes DriftWorld—the first approach to apply drift generative models to robot world modeling—by learning action-conditioned dynamics during training, enabling high-fidelity future frame sequence generation in a single forward pass. The method dramatically accelerates rollouts (over 17× faster than diffusion models) and improves policy evaluation efficiency, achieving state-of-the-art performance across multiple robotic manipulation benchmarks. Furthermore, DriftWorld functions as a high-fidelity offline simulator, with rollout scores exhibiting a 0.99 correlation with real-world task performance.

Abstract

Predictive world models enable robots to plan by imagining the outcomes of their actions, but their value for control hinges on generating many rollouts quickly. This creates a bottleneck for diffusion-based world models: multistep sampling makes each rollout expensive, limiting large-scale action search at inference time. We introduce DriftWorld, an action-conditioned world model based on drifting generative models. Rather than denoising iteratively at inference, DriftWorld learns an action-conditioned drift during training, allowing it to generate future frames from the current observation and a candidate action sequence in a single forward pass at 30+ fps, which is 17x faster on average than diffusion based baselines. We evaluate DriftWorld on standard vision-based robotic manipulation benchmarks, including Bridge-V2, RT-1, Language Table, Push-T, and Robomimic. By producing rollouts that are both accurate and fast, DriftWorld achieves state-of-the-art decision-making performance with far less inference time than diffusion-based world model baselines. Beyond online control, DriftWorld can also serve as an offline simulator for ranking real-world robot policies, with rollout-based scores correlating with ground truth at up to 0.99. These results show that drifting models are a strong fit for robot world modeling, where fast, high-quality imagination directly supports planning and policy evaluation.

Problem

• world models
• diffusion models
• robotic planning
• fast inference
• action-conditioned prediction

Innovation

• drifting generative models
• action-conditioned world model
• fast rollout generation
• robotic planning
• diffusion alternative