Data-Driven Dynamic Input Transfer for Learning Control in Multi-Agent Systems with Heterogeneous Unknown Dynamics

Learning input signals that make a dynamic system respond with a desired output is often data intensive and time consuming. It is therefore natural to ask whether, in a heterogeneous multi-agent scenario, an input signal learned by one agent can be suitably adapted and transferred to make the other agents respond with the same desired output, despite exhibiting different dynamics. In this paper, we propose a novel method to achieve this by employing a dynamic input transfer map. The method does not require any a-priori knowledge of the individual agents' dynamics. Instead, a small amount of experimental data from the source and target systems are used to estimate the transfer map. We evaluate the proposed method and compare it to existing approaches using static input transfer maps by investigating two example scenarios: (i) a simulation scenario for muscle dynamics, (ii) an experimental setting with a group of two-wheeled inverted pendulum robots and a sim-to-real motion learning problem.