SI-Diff: A Framework for Learning Search and High-Precision Insertion with a Force-Domain Diffusion Policy

RA-L 2026 & ICRA 2027

Yibo Liu, Stanko Oparnica , Simon Shewchun-Jakaitis , Guoyi (Gray) Fu,
Wang Jie, Jun Yang, Jagannathan, Anand, Tony Lo
ArXiv Paper GitHub

Overview Introduction

Training Shape

Demo Image

Cuboid Peg
35 mm × 25 mm × 60 mm
0.1 mm clearance

2 mm

5 mm

Unseen Shapes

Demo Image

Hexagonal prism
60 mm × 19 mm (flat-to-flat)
0.1 mm clearance

2 mm

5 mm

Demo Image

Pentagonal prism
60 mm × 19 mm (flat-to-flat)
0.1 mm clearance

2 mm

5 mm

Demo Image

Triangular prism
60 mm long with a side length of 34 mm
0.1 mm clearance

2 mm

5 mm

Demo Image

Cylinder
60 mm long with a diameter of 20 mm
0.1 mm clearance

2 mm

5 mm

Demo Image

Standard USB-A

2 mm

5 mm

Citation: 
@ARTICLE{11520677,
  author={Liu, Yibo and Oparnica, Stanko and Shewchun-Jakaitis, Simon and Fu, Guoyi and Wang, Jie and Yang, Jun and Jagannathan, Anand and Lo, Tony Hong-Yau},
  journal={IEEE Robotics and Automation Letters}, 
  title={SI-Diff: A Framework for Learning Search and High-Precision Insertion with a Force-Domain Diffusion Policy}, 
  year={2026},
  volume={},
  number={},
  pages={1-8},
  keywords={Modeling;Learning (artificial intelligence);Force;Training;Field-flow fractionation;Shape;Trajectory;Conferences;Forging;6-DOF;Force and Tactile Sensing;Assembly;Force Control;Diffusion Policy},
  doi={10.1109/LRA.2026.3693991}}