We propose a novel framework for learning object-centric manipulation policies in force space, decoupling the robot from the object. By directly applying forces to selected regions of the object, our method simplifies the action space, reduces unnecessary exploration, and decreases simulation overhead. This approach, trained in simulation on a small set of representative objects, captures object dynamics—such as joint configurations—allowing policies to generalize effectively to new, unseen objects.
@inproceedings{fangetal25icra,
title={FLEX: A Framework for Learning Robot-Agnostic Force-based Skills Involving Sustained Contact Object Manipulation},
author={Shijie Fang and Wenchang Gao and Shivam Goel and Christopher Thierauf and Matthias Scheutz and Jivko Sinapov},
year={2025},
booktitle={Proceedings of the 2025 IEEE International Conference on Robotics and Automation},
url={https://hrilab.tufts.edu/publications/fangetal25icra.pdf}
}