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

Abstract

This paper proposes a dataset‑aggregation approach for imitating a nonlinear model predictive control law via deep neural networks, aimed at safely allowing a robot manipulator to share its workspace with a human operator. Safety is ensured by gradually reducing the robot speed as this approaches a human operator (namely, via “speed and separation monitoring”): this condition is also imposed on the control input generated by the deep neural network via a “safety filter” based on real-time numerical optimization. The proposed method is tested experimentally 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 condition.