2nd RL-CONFORM Workshop, co-located with IROS’22

Reinforcement Learning meets HRI, Control, and Formal Methods!

TECoSA researchers Iolanda Leite. Jana Tumova. Sanne van Waveren and Alexis Linard are co-organising a 1-day workshop at IROS’22, taking place in Kyoto, Japan.

Reinforcement learning (RL) has shown remarkable achievements in applications ranging from autonomous driving, object manipulation, or beating best players in complex board-games. Different communities, including RL, human-robot interaction (HRI), control, and formal methods (FM), have proposed multiple techniques to increase safety, transparency, and robustness of RL. However, elementary problems of RL remain open: exploratory and learned policies may cause unsafe situations, lack task-robustness, or be unstable. By satisfactorily addressing these problems, RL research will have long-lasting impact and see breakthroughs on real physical systems and in human-centered environments. As an example, a collaborative mobile manipulator needs to be robust and verifiably safe around humans. This requires an integrated approach with RL to learn optimal policies for complex manipulation tasks, control techniques to ensure stability of the system, FM techniques to provide formal guarantees to ensure safety, and techniques from human-robot interaction to learn from and interact with humans.

The aim of this multidisciplinary workshop is to bring these communities together to:

1. Identify key challenges and opportunities related to safe and robust exploration, formal safety and stability guarantees of control systems, safety in physical human-robot collaborative systems;
2. Provide unique insights into how these challenges depend on the application, desired system properties, and complexity of the environment;
3. Propose new and debate existing approaches to ensure desired properties of learned policies in a wide range of domains;
4. Discuss existing and new benchmarks to accelerate safe and robust RL research;
5. Disseminate the outcomes of the workshop and publish the results as a perspectives article in one of the major robotics journals.

There will be two interactive panel sessions, one on Principles and understanding of RL algorithms and models and one on Benchmarks, implementation, and accelerating RL research.

For the full program, please see
https://rlconform-workshop.github.io/