Data-Efficient Behavior Prediction for Safe Human-Robot Collaboration

Abstract

Predicting human behavior is critical to facilitate safe and efficient human-robot collaboration (HRC) due to frequent close interactions between humans and robots. However, human behavior is difficult to predict since it is diverse and time-variant in nature. In addition, human motion data is potentially noisy due to the inevitable sensor noise. Therefore, it is expensive to collect a motion dataset that comprehensively covers all possible scenarios. The high cost of data collection leads to the scarcity of human motion data in certain scenarios, and therefore, causes difficulties in constructing robust and reliable behavior predictors. This thesis uses online adaptation (an online approach) and data augmentation (an offline approach) to deal with the data scarcity challenge in human motion prediction and intention prediction respectively. This thesis proposes a novel adaptable human motion prediction framework, RNNIK-MKF, which combines a recurrent neural network (RNN) and inverse kinematics (IK) to predict human upper limb motion. A modified Kalman filter (MKF) is applied to robustly adapt the model online to improve the prediction accuracy, where the model is learned from scarce training data. In addition, a novel training framework, iterative adversarial data augmentation (IADA), is proposed to learn safe neural network classifiers for intention prediction. The IADA uses data augmentation and expert guidance offline to augment the scarce data during the training phase and learn robust neural network models from the augmented data. The proposed RNNIK-MKF and IADA are tested on a collected human motion dataset. The experiments demonstrate that our methods can achieve more robust and accurate prediction performance comparing to existing methods.