Mini-diffuser reduces by an order of magnitude the time and memory needed to train multi-task vision-language robotic diffusion policies!

Created on June 07, 2025

2025

Mini Diffuser [1] is a game-changer for training multi-task robotic diffusion policies. We’ve slashed training time by 95% and memory by 93% while maintaining 95% performance! Train a powerful robot policy on a single GPU in just 13 hours.

Check out the code & videos: https://mini-diffuse-actor.github.io

References

[1]

Mini Diffuser: Fast Multi-task Diffusion Policy Training Using Two-level Mini-batches.

Yutong Hu, Pinhao Song, Kehan Wen, and Renaud Detry.

2025.

@misc{hu2025a,
  author = {Hu, Yutong and Song, Pinhao and Wen, Kehan and Detry, Renaud},
  title = {Mini Diffuser: Fast Multi-task Diffusion Policy Training Using Two-level Mini-batches},
  year = {2025},
  archiveprefix = {arXiv},
  eprint = {2505.09430},
  primaryclass = {cs.RO},
  selected-id = {3},
  url = {https://arxiv.org/abs/2505.09430}
}