Guidance method lets trained robot policies steer around new obstacles without retraining

Researchers led by Guangyuan Wang, with Alois Knoll among the authors, have described in an arXiv paper a way to make pretrained diffusion policies avoid obstacles they have not seen, without new demonstrations or retraining. Earlier guidance methods correct each step of a planned action sequence separately, which can fix local collisions but fails to produce a complete detour.

The new framework turns the policy's relative actions into a workspace trajectory, checks it against an obstacle model based on signed distance fields, and spreads each correction across all the action steps that led there while the policy generates its plan. Across seven obstacle scenarios in four manipulation tasks on a real robot, it completed 105 of 140 runs without collisions, against 8 of 140 for pointwise guidance on absolute actions.