Highlights from GRASP at CoRL 2022

Written by Jillian Mallon

The sixth annual Conference on Robot Learning (CoRL 2022) was held on the campus of the University of Auckland in New Zealand from December 14^th to December 18^th, 2022. The event proved to be a great opportunity for GRASP to showcase our exciting projects and papers to a wide audience of other University robotics departments as well as corporate robotics companies.

One of the 11 workshops during CoRL was co-organized by GRASP Faculty member Dr. Nadia Figueroa. The workshop titled Geometry, Physics, and Human Knowledge as Inductive Bias in Robot Learning featured speakers from subdisciplines of robot learning such as geometry, physics and human-in-the-loop learning to discuss their experience with and ideas for introducing inductive bias into robot learning models.

GRASP students also made an impressive impact on the CoRL 2022 workshops. Perception, Action and Learning Group members Edward Hu, Richard Chang, Oleh Rybkin, and GRASP faculty member Dr. Dinesh Jayaraman authored a paper titled Planning Goals for Exploration that was accepted by the Learning, Perception, and Abstraction for Long-Horizon Planning Workshop. This same paper was not only accepted by the Learning to Adapt and Improve in the Real World Workshop (RoboAdapt), but it was also highlighted in a spotlight talk and won the RoboAdapt Workshop’s Best Paper Award. RoboAdapt also accepted the paper VIP: Towards Universal Visual Reward and Representation via Value-Implicit Pre-Training that was co-authored by, among others, GRASP PhD student Jason Ma and his advisors Dr. Dinesh Jayaraman and Dr. Osbert Bastani.

The final workshop of the conference that featured the work of GRASP members was the hybrid workshop on Pre-training Robot Learning. The workshop featured virtual spotlight talks on two GRASP-affiliated papers. One of these papers was Jason Ma and Dr. Dinesh Jayaraman’s previously mentioned VIP: Towards Universal Visual Reward and Representation via Value-Implicit Pre-Training. The other, Robotic Manipulation Datasets for Offline Compositional Reinforcement Learning, was co-authored by Lifelong Machine Learning Group members Dr. Eric Eaton, Postdoctoral Researcher Cassandra Kent, PhD Student Marcel Hussing, and GRASP alumnus Jorge Mendez.

GRASPees also made significant contributions to the main program of the conference. For example, Dr. Nadia Figueroa was a co-author on a paper titled Temporal Logic Imitation: Learning Plan-Satisficing Motion Policies from Demonstrations that was accepted for both an oral presentation and poster session on December 16th. Dr. Dinesh Jayaraman served as a session chair for the conference’s Oral Session 3 on Perception on December 17th. A paper titled Online Dynamics Learning for Predictive Control with an Application to Aerial Robots, which was co-authored by ScalAR Lab members Dr. M. Ani Hsieh, CIS PhD student Tom Jiahao Zhang, and ESE PhD student Kong Yao Chee.

The final victory for GRASP at CoRL 2022 came from a paper co-authored by GRASP alumnus Kun Huang, CIS PhD student Edward Hu, and Dr. Dinesh Jayaraman titled Training Robots to Evaluate Robots: Example-Based Interactive Reward Functions for Policy Learning. After presenting the work during Oral Session 2 and Poster Session 4 on December 17th, the paper was awarded CoRL 2022’s Best Paper Award!

Congratulations go out to all of our students and faculty who represented the innovations and achievements of GRASP at CoRL 2022!