OZR Back-up mandate: Using Human Guidance in Reinforcement Learning for Multi-Robot Systems

The complexity and the distributed nature of many industrial tasks motivates the use of Multi-Robot Systems (MRS). In many applications it is unpractical to use a single robot. Also, many industrial tasks e.g. assembly require multiple robots with different capabilities to cooperate on the task. Moreover, the intrinsic redundancy of robots working in cooperative teams allows them to be more robust, besides their ability to solve a problem faster using parallelism. In general, controlling robots that can perform real-world tasks implies a big variety of challenges. Usually, the real world comes with a lot of inherent complexities e.g. the absence of perfect models. In Reinforcement Learning (RL), the agent can learn through direct interaction with the environment without the need of perfect models. This makes RL relevant for controlling robots in real-world tasks when they are hard to model. However, applying RL to robots comes also with challenges. One challenge is that it usually takes an RL agent a huge number of trials before it can learn a task, which is prohibitively expensive. Instead, the RL agent should have the capacity of making use of all available prior knowledge. In single agent settings, this has already been investigated. In our work, we will develop a novel methodology for incorporating different forms of human guidance in RL for MRS. We will have heterogeneous team of robots cooperating on an assembly task as a demonstration.