Human Motion Prediction using Semi-adaptable Neural Networks

Abstract

Human motion prediction is an important component to facilitate human robot interaction. Robots need to accurately predict human's future movement in order to safely plan its own motion trajectories and efficiently collaborate with humans. Many recent approaches predict human's movement using deep learning methods, such as recurrent neural networks. However, existing methods lack the ability to adapt to time-varying human behaviors, and many of them do not quantify uncertainties in the prediction. This paper proposes an approach that uses a semi-adaptable neural network for human motion prediction, and provides uncertainty bounds of the predictions in real time. In particular, a neural network is trained offline to represent the human motion transition model, and then recursive least square parameter adaptation algorithm (RLS-PAA) is adopted for online parameter adaptation of the neural network and for uncertainty estimation. Experiments on several human motion datasets verify that the proposed method significantly outperforms the state-of-the-art approach in terms of prediction accuracy and computation efficiency. * Equal contribution. Y. Cheng and M. Tomizuka are with the

Machine Learning Approaches for Motor Learning: A Short Review

“…Finally, RNN can also be used to learn movement embeddings, although this is not the most common approach. In the context of human-robot interaction, Cheng et al (2019) trained an RNNbased movement model offline and then adapted the last layer parameters through recursive least square errors. The goal is to adapt the robot control command to human behavior in real-time.…”

Robust nonlinear adaptation algorithms for multitask prediction networks

“…The goal of adaptation is to improve the prediction accuracy in subsequent rounds. Online adaptation is currently present in a variety of research fields like behavior prediction, 1,2,6‐8 speech recognition, 9 image recognition, 10 and machine translation 11 …”

“…The algorithm should adapt the prediction model in a way that is robust to measurement noises as well as model uncertainties. Behavior prediction is a subtopic of time series prediction, 12 which includes but is not limited to vehicle trajectory prediction during autonomous driving 13 and human‐motion prediction during human‐robot collaboration 7 . Behavior prediction of other intelligent entities is important to ensure safe and efficient operation of an autonomy.…”

“…And the proposed online adaptation algorithm can also be used on single‐task trajectory prediction. In conventional online adaptation frameworks, the online adaptation algorithm updates the prediction model according to the current single‐step feedback, although the prediction model is used to predict m ‐step ( m > 1) information in the future 1,7 . To take advantage of feedback information from more time steps in the adaptation, we introduce a ‐step feedback adaptation strategy to update the model with observations from recent steps.…”

“…Few work has studied which set of parameters (eg, which layer of the neural network model) should be adapted to achieve better results prior to this study. The common practice is to adapt the last layer of the model 1,7 or to adapt all weights of the model 17 . This article empirically compares the results of adapting different sets of parameters and suggests the best set of parameters to adapt.…”