AugInsert

Data augmentation for robust visual-force policies in contact-rich assembly (IROS 2025).

Robustness in object assembly

AugInsert studies how a dual-arm robot can finish a peg-in-hole insertion after contact begins. The policy combines visual observations, force/torque readings, and proprioception. Online augmentation expands a limited set of human demonstrations with changes in grasp pose, object geometry, appearance, camera pose, and sensor noise.

The paper evaluates these factors in simulation and finds that grasp variation is particularly challenging. It also studies the contribution of force/torque sensing to robustness. I co-authored this work with Ryan Diaz, Vivek Veeriah, and Karthik Desingh; it appeared at IROS 2025.

Read the paper.

Paper video and figures

Real-world augmentation study animation from the public AugInsert project page. See the paper for the evaluation protocol.

The AugInsert project page has the model diagram, augmentation examples, and evaluation videos. The results shown there are part of the published study; the page distinguishes simulation evaluations from its real-world augmentation study.