Teaching Robots to Predict Human Motion

摘要: Teaching a robot to predict and mimic how human moves or acts in the near future by observing series of historical movements is crucial first step human-robot interaction collaboration. In this paper, we instrument with such prediction ability leveraging recent deep learning computer vision techniques. First, our system takes images from camera as input produce corresponding skeleton based on real-time pose estimation obtained OpenPose library. Then, conditioning sequence, forecasts plausible motion through predictor, generating demonstration. Because lack high-level fidelity validation, existing forecasting algorithms suffer error accumulation inaccurate prediction. Inspired generative adversarial networks (GANs), introduce global discriminator that examines whether predicted sequence smooth realistic. Our resulting GAN model achieves superior performance state-of-the-art approaches when evaluated standard H3.6M dataset. Based model, demonstrates its replay human-like manner interacting person.