Efficient Multi-Task and Multi-Robot Transfer with Continued Learning

Ideally, robots should learn from a few demonstrations of a given task, and generalize knowledge to new, unseen tasks, and to different robots. In this paper, we focus on trajectory tracking and introduce a multi-robot, multi-task transfer learning framework that allows a target system to complete a target task by learning from a few demonstrations of a source task on a source system. The proposed multi-robot transfer learning framework is based on a combined L1 adaptive control and iterative learning control approach. The key idea is that the adaptive controller forces dynamically different systems to behave as the specified reference model. The proposed multi-task transfer learning framework uses theoretical control results (e.g. the concept of vector relative degree) to learn a map from desired trajectories to the inputs that make the system track these trajectories. The learned map is then used to calculate the inputs for a new, unseen trajectory. Experimental results using two different quadrotors show that, using the proposed framework, it is possible to significantly reduce the tracking error of a target trajectory on the target system when information from a single source trajectory learned on the source system is used.