Optimizing and Predicting Human-Symbiotic Robot Trajectory

This paper introduces human-symbiotic trajectory prediction (HSTP) for mobile robots that interact with humans in close proximity. HSTP predicts an optimized motion that is friendly to both humans and robots, considering both the closeness of the prediction to human motion and the smoothness of robot control. The optimized trajectory is learned by a neural network using a large-scale human motion dataset to embed its general motion feature. Then, this network is used to predict the optimal future robot trajectory using several frames of motion. The proposed model was validated by simulating a human coincidence robot (HCR) controlled based on predictions from human motion data measured using an RGB-D sensor. The proposed HSTP model significantly reduced the tracking error by more than 70% under conditions with the largest differences. Furthermore, the model compensates for sensor information delays of up to 1.0 $s$ through its predictive capabilities.