Diffusion Policy Lets Robots Open Heavy Doors and Pass Through with Dual Arms

Opening heavy, self-closing doors—especially those requiring pulling—has been a long-standing challenge in robotics. Humans naturally use both arms dexterously: rotating the handle, widening the gap, holding the door, switching arms, and moving through while maintaining clearance. Traditional robotic approaches rely on state machines with manually defined stages that lack robustness to real-world variability. Enter diffusion policy: researchers from UMass (Shangqun Yu, Matthew En, Daniel Wu, Sangjun Park, Ziyi Zhou, Seyed Fakoorian, and Donghyun Kim) trained a single end-to-end diffusion-based visuomotor control policy that simultaneously coordinates a nonholonomic mobile base and dual arms for the complete door opening and passing task.

The result is a policy that not only achieves a high success rate in opening and traversing damped pull doors but also demonstrates strong robustness to external disturbances—capabilities difficult to realize with classical methods. This work advances imitation learning for long-horizon tasks requiring tight coordination between manipulation and locomotion, potentially enabling robots to operate in human-centric environments with heavy doors (e.g., warehouses, hospitals, offices). The paper is available on arXiv (2605.15352) and represents a significant step toward more autonomous mobile manipulators.