Focusing on task-critical object parts raises DINOBot's imitation success by 13.8%

Researchers led by Denis Hadjivelichkov have described two changes to DINOBot, a framework that teaches robots interactions between objects by imitation, in a paper accepted at ICAR 2025.

First, they add RANSAC outlier rejection to the alignment step to limit the effect of noisy depth data and false feature matches. Second, they focus semantic matching on the regions of objects where the interaction happens, so the learned representation favours the features that decide whether a manipulation succeeds.

In real-world picking, insertion, shaking and pouring tasks, task success rose by 13.8% overall compared with the baseline, and alignment converged faster, the authors report.