Reproduce “dWorldEval: Scalable Robotic Policy Evaluation via Discrete Diffusion World Model” on OpenResearch

About this paper

Yaxuan Li, Zhongyi Zhou, Yefei Chen, Yaokai Xue, Yichen Zhu

dWorldEval introduces a discrete diffusion world model for scalable robotic policy evaluation, which enhances action controllability and spatiotemporal consistency by treating actions as primary tokens. The model demonstrates high correlation with real-world policy performance across diverse tasks and platforms, allowing for reliable and automated assessment of robotic policies.

Key insights

• Integrating actions as primary tokens within a unified discrete diffusion framework, trained from scratch on robotic data, significantly improves action controllability and prevents the hallucination of successes often seen in models adapting video generation architectures.
• The use of sparse keyframe memory is crucial for maintaining spatiotemporal consistency and bounding cumulative visual drift over long imagined horizons in robotic world models.
• Jointly predicting a discrete progress token alongside visual outcomes provides an effective and automatic mechanism for task success detection, aligning assessment directly with the generated visual trajectory.

Summary by alphaXiv.