New CoRL paper: FOCI Policy: Focus on Object-Centric Interactions for Relational Manipulation!

FOCI Policy: Focus on Object-Centric Interactions for Relational Manipulation 🤖 [1]

Most manipulation policies learn observation → action mappings and pay for it in demonstrations. We took another look at how far the opposite bet goes: make object-centric relative SE(3) motion the unit of learning, narrowed to the short, tightly-constrained interaction segments that change-point detection pulls out of each demo. How much of what action-centric policies learn from scale is really an invariance we could just build in?

Check out the code & videos: https://fitz0401.github.io/foci-page/

References

CoRL 2026

[1]

FOCI Policy: Focus on Object-Centric Interactions for Relational Manipulation Policies.

Ze Fu, Pinhao Song, Yutong Hu, and Renaud Detry.

In Conference on Robot Learning, 2026.

@inproceedings{fu2026c,
  author = {Fu, Ze and Song, Pinhao and Hu, Yutong and Detry, Renaud},
  title = {FOCI Policy: Focus on Object-Centric Interactions for Relational Manipulation Policies},
  year = {2026},
  archiveprefix = {arXiv},
  booktitle = {Conference on Robot Learning},
  primaryclass = {cs.RO},
  url = {https://arxiv.org/abs/2609.08743}
}