Concurrent Skill Composition Using Ensemble of Primitive Skills

One of the key characteristics of an open-ended cumulative learning agent is that it should use the knowledge gained from prior learning to solve future tasks. That characteristic is especially essential in robotics, as learning every perception-action skill from scratch is not only time consuming but may not always be feasible. In the case of reinforcement learning, this learned knowledge is called a policy. The lifelong learning agent should treat the policies of learned tasks as building blocks to solve those future tasks. One of the categorizations of tasks is based on its composition, ranging from primitive tasks to compound tasks that are either a sequential or concurrent combination of primitive tasks. Thus, the agent needs to be able to combine the policies of the primitive tasks to solve compound tasks, which are then added to its knowledge base. Inspired by modular neural networks, we propose an approach to compose policies for compound tasks that are concurrent combinations of disjoint tasks. Furthermore, we hypothesize that learning in a specialized environment leads to more efficient learning; hence, we create scaffolded environments for the robot to learn primitive skills for our mobile robot-based experiments. We then show how the agent can combine those primitive skills to learn solutions for compound tasks. That reduces the overall training time of multiple skills and creates a versatile agent that can mix and match the skills.