Augmentation Enables One-Shot Generalization In Learning From Demonstration for Contact-Rich Manipulation
Summary
We introduce a Learning from Demonstration (LfD) approach for contact-rich manipulation tasks with articulated mechanisms.
We showcase that the extracted policy from a single augmented human demonstration generalizes to different mechanisms of the same type and achieves high success rates even in a changing environment.
Section III. A: Augmentation for Contact
Based on an opening demonstration, our approach uncovers the underlying contact information through augmentation.
Section III. B: Augmentation for Vision
Based on a demonstration of grasping, our approach collects self-supervised images to train an object pose estimator for grasping.
Videos For Experiments (Section VI)
Here are the videos for experimental results with a real robot. These experimental results on complex mechanisms with multi-DOF demonstrate that our approach can reliably accomplish the task in a changing environment.
Section VI-A: Opening Latch Locks
Our approach robustly opens locks against environmental changes. It also generalizes to different locks of the same type. The key to such substantial robustness and generalization just from a single demonstration is integrating augmentation into Learning from Demonstration to reveal additional helpful information about the environment!
Our approach achieves a 100% success rate for the first two locks and 70% for the last lock, in total a 90% for all locks.
Section VI-B: Experiment on Drawers
We show here a single augmented demonstration with a drawer suffices to open another drawer in a very robust way.
The 100% success rates for both drawers confirm that our approach can not only operate various mechanisms robustly – from a single augmented demonstration – but also that it yields behavior that generalizes to unseen object instances of the same type.
Section VI-C: Opening Chain Locks
We performed additional experiments with chain locks to exemplify that our approach can apply to various mechanisms.
Overall, our approach succeeds 100% for chain lock 1 and 90% times for chain lock 2 and 3, respectively.
We gratefully acknowledge funding by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy – EXC 2002/1 “Science of Intelligence” – project number 390523135.