Deep Adaptive Indirect Herding of Multiple Target Agents with Unknown Interaction Dynamics

The herding problem has received growing interest in the robotics and controls community in recent years. In particular, indirect herding is an abstraction for many potential applications in fields such as wildlife management, crowd control, traffic management, and environment cleanups. Existing works in indirect herding, however, have not taken advantage of recent advances in the field of adaptive control, which have allowed for the development of adaptive controllers using Lyapunov-based deep neural networks (Lb-DNNs). These results, however, are only applicable for systems that are directly controlled, and not for indirect control problems such as the herding problem. This paper develops a novel approach to address the indirect herding problem using an Lb-DNN adaptive backstepping design. The Lb-DNN adaptive backstepping controller enables the herding agent to learn the interaction dynamics and adaptively herd the target agents in real-time, using actual interactions during task-execution. A Lyapunov-based switched systems analysis is used to develop sufficient dwell-time conditions which guarantee exponential convergence of all states to an ultimate bound. Simulations are provided to demonstrate the performance of the developed Lb-DNN adaptive backstepping controller.