Safely Imitating Predictive Control Policies for Real-Time Human-Aware Manipulator Motion Planning: A Dataset Aggregation Approach

This paper proposes a dataset-aggregation approach for imitating a nonlinear model predictive control law via deep neural networks, to safely allow a robot manipulator to share its workspace with a human operator. As the robot approaches the human, its speed is gradually reduced using the “speed and separation monitoring” framework. Specific time-varying upper bounds are explicitly imposed on the control input generated by the deep neural network through a “safety filter” based on real-time numerical optimization. The proposed method is experimentally tested on a UR5 manipulator, comparing the performance of different neural network structures and types of training. As a result, it is shown that the dataset-aggregation approach provides better performance with respect to a “naive” approach to training, and that the presence of the safety filter is indeed needed to avoid the violation of the speed-and-separation-monitoring constraints.