# Set-Supervised Diffusion Policy (SDP): Learning Action-Chunking Diffusion through Corrections

> SDP is a method for training action-chunking diffusion policies from human-in-the-loop corrections. Instead of only imitating the human's corrective action, SDP uses both the corrective (positive) and the robot's undesired (negative) action-chunks to construct a set of desired action-chunks, and trains the diffusion policy to align with that set. Published at Robotics: Science and Systems (RSS 2026).

- Authors: Zhaoting Li, Gang Chen, Javier Alonso-Mora, Cosimo Della Santina (Delft University of Technology); Jens Kober (University of Stuttgart)
- Venue: Proceedings of Robotics: Science and Systems XXII (RSS 2026), Sydney, Australia, July 2026
- DOI: 10.15607/RSS.2026.XXII.080
- arXiv: 2606.01865
- Topics: diffusion policy, action chunking, interactive imitation learning, human-in-the-loop corrections, contrastive / negative supervision, behavior cloning, distributional shift, robotic manipulation

## Links

- [Paper (RSS 2026 proceedings)](https://www.roboticsproceedings.org/rss22/p080.html)
- [Paper (PDF, RSS proceedings)](https://www.roboticsproceedings.org/rss22/p080.pdf)
- [DOI](https://doi.org/10.15607/RSS.2026.XXII.080)
- [arXiv preprint](https://arxiv.org/abs/2606.01865)
- [Code (GitHub)](https://github.com/ZhaotingLi/Set_Supervised_DP)
- [Project page](https://set-supervised-diffusion-policy.github.io/)

## Abstract

Diffusion policies have recently emerged as a powerful framework for robotic manipulation. However, like other behavior cloning methods, they remain vulnerable to distributional shift, often requiring human-in-the-loop interventions to correct failures during deployment. These interactions naturally provide paired supervision in the form of the robot's undesired actions and the human teacher's corrective actions. Yet existing data aggregation pipelines and standard behavior cloning losses largely ignore this negative signal from undesired actions, leading to overfitting to teacher's actions and an increasing reliance on costly expert data.

To address this limitation, we propose Set-Supervised Diffusion Policy (SDP), a novel learning framework that utilizes contrastive action-chunk data to train diffusion policies from human corrections. From paired positive and negative action-chunks, SDP constructs a set of desired action-chunks and designs a training pipeline that encourages the diffusion policy to align with the set. Through extensive experiments across multiple robotic manipulation tasks, we demonstrate that SDP consistently improves policy performance, with particularly strong gains in robustness to noisy data. Moreover, SDP induces high-quality aggregated datasets, enabling more efficient and reliable policy learning from human-in-the-loop corrections.

## Real-robot experiments

- Insert-T: contact-rich, multi-modal pushing task to align and insert a T-shaped object into a U-shaped object. SDP consistently improves over Diffusion Policy (DP), with the largest gains on harder initial states.
- Round-table assembly: long-horizon assembly where small errors compound across stages. Corrections concentrate around bottleneck stages (e.g., inserting parts); SDP makes better use of this feedback than DP, giving stronger stage-wise and overall success.
- Human corrections are given with a 6D space mouse during deployment; the robot's attempted action is treated as negative and the human correction as positive.

## Citation

```bibtex
@INPROCEEDINGS{LiZ-RSS-26,
    AUTHOR    = {Zhaoting Li AND Gang Chen AND Javier Alonso-Mora AND Cosimo Della Santina AND Jens Kober},
    TITLE     = {{Set-Supervised Diffusion Policy: Learning Action-Chunking Diffusion through Corrections}},
    BOOKTITLE = {Proceedings of Robotics: Science and Systems},
    YEAR      = {2026},
    ADDRESS   = {Sydney, Australia},
    MONTH     = {July},
    DOI       = {10.15607/RSS.2026.XXII.080}
}
```