Active Training Trajectory Generation for Inverse Dynamics Model Learning with Deep Neural Networks

Inverse dynamics models have been used in robot control algorithms to realize a desired motion or to enhance a robot's performance. As robot dynamics and their operating environments become more complex, there is a growing trend of learning uncertain or unknown dynamics from data. While techniques such as deep neural networks (DNNs) have been successfully used to learn inverse dynamics, it is usually implicitly assumed that the learning modules are trained on sufficiently rich datasets. In practical implementations, this assumption typically results in a trial-and-error training process, which can be inefficient or unsafe for robot applications. In this paper, we present an active trajectory generation framework that allows us to systematically design informative trajectories for training DNN inverse dynamics modules. In particular, we introduce an episode-based algorithm that integrates a spline trajectory optimization approach with DNN active learning for efficient data collection. We consider different DNN uncertainty estimation techniques and active learning heuristics in our work and illustrate the proposed active training trajectory generation approach in simulation. We show that the proposed active training trajectory generation outperforms adhoc, intuitive training approaches.