Consistent Iterative Denoising for Robot Manipulation
Keywords: robot manipulation, consistent iterative denoising, diffusion model, imitation learning
Abstract: Robot manipulation in complex scenarios usually involves multiple successful actions, which requires generative models to estimate the distribution of various successful actions.
In recent years, the diffusion model has been widely studied in many robot manipulation tasks.
However, the diffusion model experiences inconsistent noise supervision across various action labels and denoising timesteps, which compromises accurate action prediction.
On the one hand, CIDM designs new noise supervision to avoid interference between different successful actions, leading to consistent denoising directions.
On the other hand, CIDM unifies all denoising timesteps, avoiding inconsistent predictions of the diffusion model over different timesteps.
Moreover, we also designed a novel radial loss to make the model focus on denoising results rather than iterative process routes.
Our method achieves a new state-of-the-art performance on RLBench with the highest success rate of 82.3\% on a multi-view setup and 83.9\% on a single-view setup.
Primary Area: applications to robotics, autonomy, planning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4510
Loading