Scaling Robot Policy Evaluation via Discrete Diffusion World Model

Yaxuan Li, Junjie Wen, Zhongyi Zhou, Yefei Chen, Chaomin Shen, Yaxin Peng, Yichen Zhu
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:71168-71186, 2026.

Abstract

Evaluating generalist robot manipulation policies is costly and difficult to scale in the real world. While emerging world models (e.g., WorldEval, Ctrl-World) offer a promising alternative, the reliability of such evaluation remains a critical bottleneck. Specifically, their visual predictions can undermine policy assessment by "self-correcting" failures into false positives or yielding artifacts under out-of-distribution controls. Even with failure-enriched data, current architectures struggle to capture action-causal dynamics, as they typically treat actions as passive conditions rather than causal drivers. To address this, we propose dWorldEval, an action-centric discrete-diffusion world model that maps visual observations, language instructions, and action chunks into a shared unified token space and denoises them with a single self-attention backbone where actions function as first-class tokens. To realize reliable policy-world interaction, dWorldEval introduces a sparse keyframe memory that anchors global scene state while preserving fine-grained multi-view interaction cues, and leverages Progress-as-text to jointly generate future observations and success indicators. Extensive experiments on LIBERO, RoboTwin, and real-robot tasks demonstrate that dWorldEval significantly outperforms video diffusion baselines in action controllability, stabilizes long-horizon multi-view rollouts, enabling accurate policy ranking via automatic success estimation.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-li26hi, title = {Scaling Robot Policy Evaluation via Discrete Diffusion World Model}, author = {Li, Yaxuan and Wen, Junjie and Zhou, Zhongyi and Chen, Yefei and Shen, Chaomin and Peng, Yaxin and Zhu, Yichen}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {71168--71186}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/li26hi/li26hi.pdf}, url = {https://proceedings.mlr.press/v306/li26hi.html}, abstract = {Evaluating generalist robot manipulation policies is costly and difficult to scale in the real world. While emerging world models (e.g., WorldEval, Ctrl-World) offer a promising alternative, the reliability of such evaluation remains a critical bottleneck. Specifically, their visual predictions can undermine policy assessment by "self-correcting" failures into false positives or yielding artifacts under out-of-distribution controls. Even with failure-enriched data, current architectures struggle to capture action-causal dynamics, as they typically treat actions as passive conditions rather than causal drivers. To address this, we propose dWorldEval, an action-centric discrete-diffusion world model that maps visual observations, language instructions, and action chunks into a shared unified token space and denoises them with a single self-attention backbone where actions function as first-class tokens. To realize reliable policy-world interaction, dWorldEval introduces a sparse keyframe memory that anchors global scene state while preserving fine-grained multi-view interaction cues, and leverages Progress-as-text to jointly generate future observations and success indicators. Extensive experiments on LIBERO, RoboTwin, and real-robot tasks demonstrate that dWorldEval significantly outperforms video diffusion baselines in action controllability, stabilizes long-horizon multi-view rollouts, enabling accurate policy ranking via automatic success estimation.} }
Endnote
%0 Conference Paper %T Scaling Robot Policy Evaluation via Discrete Diffusion World Model %A Yaxuan Li %A Junjie Wen %A Zhongyi Zhou %A Yefei Chen %A Chaomin Shen %A Yaxin Peng %A Yichen Zhu %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-li26hi %I PMLR %P 71168--71186 %U https://proceedings.mlr.press/v306/li26hi.html %V 306 %X Evaluating generalist robot manipulation policies is costly and difficult to scale in the real world. While emerging world models (e.g., WorldEval, Ctrl-World) offer a promising alternative, the reliability of such evaluation remains a critical bottleneck. Specifically, their visual predictions can undermine policy assessment by "self-correcting" failures into false positives or yielding artifacts under out-of-distribution controls. Even with failure-enriched data, current architectures struggle to capture action-causal dynamics, as they typically treat actions as passive conditions rather than causal drivers. To address this, we propose dWorldEval, an action-centric discrete-diffusion world model that maps visual observations, language instructions, and action chunks into a shared unified token space and denoises them with a single self-attention backbone where actions function as first-class tokens. To realize reliable policy-world interaction, dWorldEval introduces a sparse keyframe memory that anchors global scene state while preserving fine-grained multi-view interaction cues, and leverages Progress-as-text to jointly generate future observations and success indicators. Extensive experiments on LIBERO, RoboTwin, and real-robot tasks demonstrate that dWorldEval significantly outperforms video diffusion baselines in action controllability, stabilizes long-horizon multi-view rollouts, enabling accurate policy ranking via automatic success estimation.
APA
Li, Y., Wen, J., Zhou, Z., Chen, Y., Shen, C., Peng, Y. & Zhu, Y.. (2026). Scaling Robot Policy Evaluation via Discrete Diffusion World Model. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:71168-71186 Available from https://proceedings.mlr.press/v306/li26hi.html.

Related Material