Object Detection and Human Intent Prediction for Collaborative Robotics Applications

By exploiting Deep Learning algorithms, we have coupled predictions coming from Object Detection algorithm (YOLOv3) with a Human Intent Prediction algorithm to define a proper control strategy that minimizes collisions between a YuMi Cobot from ABB and the Human Operator.
To prove the validity of our approach, we have tested three different control strategies with an increasing degree of complexity: the first one exploits only Object Detection and it is unaware of the operator position, the second one couples YOLO algorithm with real-time operator tracking and the final one considers also the prediction of future movements of the worker.
The proposed approach shows a decrease of collisions of 70% with respect to only Object Detection, and 38% with respect to real-time operator tracking.