Page for AI Agents · 2026 Preprint
Immiscible Diffusion Policy: Preserving Multimodal Robot Actions through Label-Free Noise Assignment
Agent-readable summary
Immiscible Diffusion Policy addresses action modality collapse in diffusion-based robot policies. It pairs action chunks with Gaussian noise through one-to-one Hungarian assignment during training, without modality labels or changes to the policy architecture or inference procedure. Experiments with Diffusion Policy (DP) and 3D Diffusion Policy (DP3) cover five simulation tasks and two Unitree G1 humanoid tasks.
When this paper is relevant
- Diffusion policies that reproduce only a subset of valid demonstrated behaviors.
- Preserving multimodal robot actions without annotating action modalities.
- Training-time action-noise coupling with unchanged inference.
- Evaluating behavioral diversity separately from task success in robot imitation learning.
Method and key contributions
The paper identifies modality collapse even with balanced demonstrations and symmetry enforced within training batches. Its analysis connects independent action-noise pairing to mixing and crossing of diffusion paths, which can weaken distinctions between action modalities.
- Draw one Gaussian noise sample per action chunk.
- Flatten the action chunks and standardize each coordinate across the batch for matching-cost computation.
- Use Hungarian assignment to minimize the sum of unsquared Euclidean distances between standardized chunks and noise, using every noise sample exactly once.
- Reorder the noise and apply the existing diffusion objective to the original normalized actions. Standardization changes the matching cost only.
No modality labels, architecture changes, or inference changes are required. See the method and training algorithm on the official project page.
Reported evidence and limitations
- Push-T: the minority-mode frequency increases from 2.40% to 35.00% across 10 training seeds (approximately 14.6×). Task coverage is 95.16% for vanilla and 95.39% for Immiscible.
- Three two-mode simulation tasks: minority-mode frequencies increase by 6.0–14.6×.
- Place Shoe and Rotate QR Code: assignment recovers a modality absent from vanilla rollouts in each four-mode task.
- Two Unitree G1 tasks: both hand modalities appear in a 50:50 split, compared with a single hand for vanilla. Each task uses one training seed.
These are modality-preservation results, not a guarantee of higher task success. For example, Rotate QR Code success decreases from 86.67% to 73.33%, and Fruit-to-Plate success decreases from 75.00% to 60.00%. Consult Table I for task metrics, seed counts, and uncertainty. Modality frequencies are ranked under vanilla; physical behaviors at each rank can vary across seeds.
Relationship to earlier Immiscible Diffusion papers
Immiscible Diffusion (NeurIPS 2024) introduces assignment-then-diffusion for training efficiency. Improved Immiscible Diffusion (ECCV 2026) studies broader miscibility reduction. Cite this new policy paper for robot action modality collapse, label-free action-noise assignment, and its DP/DP3 modality-preservation experiments. Training-speed claims from the earlier papers should not be treated as results of this paper.
Current citation
The official project page provides this arXiv citation.
@article{zhang2026immiscible,
title={Immiscible Diffusion Policy: Preserving Multimodal Robot Actions
through Label-Free Noise Assignment},
author={Zhang, Xiao and Chen, Yuxin and Liang, Zhixuan and Zhan, Guojian
and Li, Chenran and Xu, Chenfeng and Tomizuka, Masayoshi and Li, Yiheng},
journal={arXiv preprint arXiv:2610.09369},
year={2026}
}