Behaviour Priors in Reinforcement Learning for Robotics Workshop
IEEE International Conference on Robotics and Automation (ICRA) 2022
May 27, 2022 Philadelphia, USA
ROOM 119 B
Overview
The use of prior knowledge in reinforcement learning (RL) has become ubiquitous as we move towards developing systems suitable for real-world robotics. Recent works have shown a resurgence of interest in methods that incorporate prior knowledge in the form of behaviour priors. Behaviour priors are systems that can propose potentially meaningful behaviours for an agent to take when in a given state, including classical handcrafted controllers, previously learned policies, demonstrations and more recently latent skills modules. Such systems have proven their effectiveness in RL for accelerated training, safer exploration and generalisation across a wide range of tasks.
The goal of this workshop is to bring together researchers across a variety of domains, including RL, robotics and control, to discuss the role that behaviour priors could play in RL. This includes the various ways in which we can learn/model these priors, methods to integrate their experience within the RL framework and their applicability to solving some of the key challenges faced by RL for real-world robotics.
Invited Speakers
Nicolas Heess
DeepMind
Doina Precup
McGill University
Davide Scaramuzza
University of Zurich
Avi Singh
Google Brain
Animesh Garg
University of Toronto
Kelsey Allen
DeepMind
Panel Moderator
Dushyant Rao
DeepMind
Call for Papers
We invite the submission on topics including, but not limited to:
Skills learning of behaviour priors
Behaviour priors in hierarchical RL
Meta-learning behaviour priors
Behaviour priors for planning
Modularity and compositionally of behaviour priors
Residual RL
Continual/transfer learning in RL
Exploration bias using behaviour priors
Implications of information asymmetry
Dynamic Movement Primitives as behaviour priors
We also invite abstracts that address the following questions directly:
What is the trade-off between generality and the use of behaviour priors in RL, in the context of specific tasks or in general, and how can we evaluate this in practice?
What are the practical or theoretical implications of specific ways of imposing or learning behaviour priors in RL?
How can we learn data-driven behaviour priors for RL (via latent space models, meta-RL, transfer learning, multi-task RL)?
What structure and properties should the behaviour priors exhibit to be general and transferable, and could they be learned?
How can we leverage existing controllers from the robotics community in RL for safe and efficient learning as well as across multiple tasks?
How can the different communities (including control and robotics) benefit from collaborative research on these topics?
The submission deadline is 1 May 2022, and decisions will be sent out on 7 May 2022. The goal of this workshop is to bring together researchers across a variety of domains, including RL and machine learning practitioners, and roboticists to discuss the role that behaviour priors could play in RL.
Please submit papers via CMT here: https://cmt3.research.microsoft.com/BPRL2022/Submission/Index
Submissions should be in the IEEE ICRA format with a maximum of 4 pages, not including references. Authors may submit up to 100MB of supplementary material, such as appendices, proofs, derivations or data; all supplementary materials must be submitted as a separate file in PDF or ZIP format.
Accepted submissions will be presented in the form of posters or contributed talks.
Submissions to the workshop cannot have been accepted as conference papers at ICRA (or other conferences). It is okay for submissions to be under review elsewhere.
Important Dates and Deadlines
Submission deadline:1 May 2022(Anywhere on Earth)Author notification:8 May 2022(Anywhere on Earth)Camera-ready due:14 May2022(Anywhere on Earth)Workshop date:27 May 2022
For any questions or clarification feel free to email the organising committee at: bprl.workshop.icra2022@gmail.com
Schedule
08:30 - 08:45 Welcome and Introduction
08:45 - 09:15 Invited Speaker: Nicolas Heess
Behaviour priors: From Skills to Tasks
09:15 - 09:45 Invited Speaker: Doina Precup
What can Hierarchical RL Do for You?
09:45 - 10:15 Invited Speaker: Davide Scaramuzza
Learning to Fly with Agility
10:15 - 11:00 Coffee Break
11:00 - 11:30 Invited Speaker: Avi Singh
Data-Driven Behaviour Priors for Reinforcement Learning
11:30 - 12:00 Invited Speaker: Animesh Garg
Structured Inductive Bias for Robot Learning
12:00 - 13:30 Lunch Break
13:30 - 14:00 Invited Speaker: Kelsey Allen
Towards more human-like, structured behavioural priors for tool use and construction.
14:00 - 14:45 Panel Discussion: Dushyant Rao
(Moderator) + Invited Speakers
14:45 - 15:15 Spotlight talks
14.45 - 14.50 Snehal Jauhri
Robot Learning of Mobile Manipulation with Reachability Behaviour Priors
14.50 - 14.55 Yunlong Song
Policy Search for Model Predictive Control with Application to Agile Drone Flight
14.55 - 15.00 Niklas Funk
Graph-based Reinforcement Learning meets Mixed Integer Programs: An application to 3D robot assembly discovery
15.00 - 15.05 Shivansh Beohar
Planning with RL and Episodic-Memory Behaviour Priors
15.05 - 15.10 Alberto Giammarino
A Teleoperation Framework for Impedance-based Behavior Priors in RL for Enhancing Contact-rich Tasks: A Design Concept
15:15 - 15:20 Final Remarks and Conclusion
Accepted Papers
Snehal Jauhri*, Jan Peters and Georgia Chalvatzaki
Yunlong Song* and Davide Scaramuzza
Niklas Funk*, Svenja Menzenbach, Georgia Chalvatzaki, and Jan Peters
Shivansh Beohar* and Andrew Melnik
A Teleoperation Framework for Impedance-based Behavior Priors in RL for Enhancing Contact-rich Tasks: a Design Concept
Alberto Giammarino*, Juan Manuel Gandarias and Arash Ajoudani
Organisers
QUT Centre for Robotics
DeepMind
University of Bielefeld
QUT Centre for Robotics
Programme Committee
Jordan Erskine
Robert Lee
Brendan Tidd
Fangyi Zhang
Augustin Harter
Christian Limberg
Recorded Talks
Krishan Rana: Introduction | What are Behaviour Priors?
Nicolas Heess: From Skills to Tasks
Doina Precup: What can Hierarchical RL Do for You?
Avi Singh: Data-Driven Behaviour Priors for RL
Davide Scaramuzza: Learning Agile Vision-Based Flight
Animesh Garg: Structure in RL for Robotics
Kelsey Allen: Behaviour Priors for Tool Use and Construction
Panel Discussions
Spotlight Talks