STEP: Warm-Started Visuomotor Policies with Spatiotemporal Consistency Prediction

Jinhao Li, Yuxuan Cong, Yingqiao Wang, Hao Xia, Shan Huang, Yijia Zhang, Ningyi Xu, Guohao Dai
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:66501-66525, 2026.

Abstract

Diffusion policies have recently emerged as a powerful paradigm for visuomotor control in robotic manipulation due to their ability to model the distribution of action sequences and capture multimodality. However, iterative denoising leads to substantial inference latency, limiting control frequency in real-time closed-loop systems. Existing acceleration methods either reduce sampling steps, bypass diffusion through direct prediction, or reuse past actions, but often struggle to jointly preserve action quality and achieve consistently low latency. In this work, we propose STEP, a lightweight spatiotemporal consistency prediction mechanism to construct high-quality warm-start actions that are both distributionally close to the target action and temporally consistent, without compromising the generative capability of the original diffusion policy. Then, we propose a velocity-aware perturbation injection mechanism that adaptively modulates actuation excitation based on temporal action variation to prevent execution stall especially for real-world tasks. We further provide a theoretical analysis showing that the proposed prediction induces a locally contractive mapping, ensuring convergence of action errors during diffusion refinement. Extensive evaluations on nine simulated benchmarks and two real-world tasks show that STEP with 2 steps can achieve an average 21.6% and 27.5% higher success rate than BRIDGER and DDIM on the RoboMimic benchmark and real-world tasks, respectively. The code is available at https://github.com/Kimho666/STEP.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-li26aa, title = {{STEP}: Warm-Started Visuomotor Policies with Spatiotemporal Consistency Prediction}, author = {Li, Jinhao and Cong, Yuxuan and Wang, Yingqiao and Xia, Hao and Huang, Shan and Zhang, Yijia and Xu, Ningyi and Dai, Guohao}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {66501--66525}, 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/li26aa/li26aa.pdf}, url = {https://proceedings.mlr.press/v306/li26aa.html}, abstract = {Diffusion policies have recently emerged as a powerful paradigm for visuomotor control in robotic manipulation due to their ability to model the distribution of action sequences and capture multimodality. However, iterative denoising leads to substantial inference latency, limiting control frequency in real-time closed-loop systems. Existing acceleration methods either reduce sampling steps, bypass diffusion through direct prediction, or reuse past actions, but often struggle to jointly preserve action quality and achieve consistently low latency. In this work, we propose STEP, a lightweight spatiotemporal consistency prediction mechanism to construct high-quality warm-start actions that are both distributionally close to the target action and temporally consistent, without compromising the generative capability of the original diffusion policy. Then, we propose a velocity-aware perturbation injection mechanism that adaptively modulates actuation excitation based on temporal action variation to prevent execution stall especially for real-world tasks. We further provide a theoretical analysis showing that the proposed prediction induces a locally contractive mapping, ensuring convergence of action errors during diffusion refinement. Extensive evaluations on nine simulated benchmarks and two real-world tasks show that STEP with 2 steps can achieve an average 21.6% and 27.5% higher success rate than BRIDGER and DDIM on the RoboMimic benchmark and real-world tasks, respectively. The code is available at https://github.com/Kimho666/STEP.} }
Endnote
%0 Conference Paper %T STEP: Warm-Started Visuomotor Policies with Spatiotemporal Consistency Prediction %A Jinhao Li %A Yuxuan Cong %A Yingqiao Wang %A Hao Xia %A Shan Huang %A Yijia Zhang %A Ningyi Xu %A Guohao Dai %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-li26aa %I PMLR %P 66501--66525 %U https://proceedings.mlr.press/v306/li26aa.html %V 306 %X Diffusion policies have recently emerged as a powerful paradigm for visuomotor control in robotic manipulation due to their ability to model the distribution of action sequences and capture multimodality. However, iterative denoising leads to substantial inference latency, limiting control frequency in real-time closed-loop systems. Existing acceleration methods either reduce sampling steps, bypass diffusion through direct prediction, or reuse past actions, but often struggle to jointly preserve action quality and achieve consistently low latency. In this work, we propose STEP, a lightweight spatiotemporal consistency prediction mechanism to construct high-quality warm-start actions that are both distributionally close to the target action and temporally consistent, without compromising the generative capability of the original diffusion policy. Then, we propose a velocity-aware perturbation injection mechanism that adaptively modulates actuation excitation based on temporal action variation to prevent execution stall especially for real-world tasks. We further provide a theoretical analysis showing that the proposed prediction induces a locally contractive mapping, ensuring convergence of action errors during diffusion refinement. Extensive evaluations on nine simulated benchmarks and two real-world tasks show that STEP with 2 steps can achieve an average 21.6% and 27.5% higher success rate than BRIDGER and DDIM on the RoboMimic benchmark and real-world tasks, respectively. The code is available at https://github.com/Kimho666/STEP.
APA
Li, J., Cong, Y., Wang, Y., Xia, H., Huang, S., Zhang, Y., Xu, N. & Dai, G.. (2026). STEP: Warm-Started Visuomotor Policies with Spatiotemporal Consistency Prediction. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:66501-66525 Available from https://proceedings.mlr.press/v306/li26aa.html.

Related Material