Wei Mao is a first year PhD student at Australian National University (ANU) under the supervision of Miaomiao Liu. This is a joint work with Mathieu Salzmann from EPFL and Hongdong Li from ANU. He speaks to us ahead of his oral and poster this afternoon. The work is about human motion prediction , which means generating where a human is going to move in the future after you have seen their movement in the past. Human motion prediction is key to the success of applications such as self-driving cars, human-robot interaction and social robots. Wei asks us to imagine that we want a robot to collaborate with a human worker . If the robot can’t predict the human movement, then the direction of the robot will always be delayed compared to the action of the worker. To collaborate with the worker seamlessly, the human robot needs to prepare for the future movement of that worker. One key insight of the work is that the past trajectory of human joints is closely related to the future trajectory. That’s why Wei says they first encode the past trajectory trying to find a relation between the past sequence and the future sequence. Also, the movement of one joint is closely related to that of other joints, so they leverage the graph convolutional networks to capture the joint trajectory dependencies. The challenge is that humans all move quite differently from each other and predicting the future is difficult, but Wei says their method is simpler, smaller, faster and more effective than previous work: “Previous work used recurrent neural networks (RNNs). Using RNNs is a natural choice for handling sequence data, but there are a few drawbacks. The past information will be forgotten after a long propagation because you 10 Oral Presentations DA I L Y Learning Trajectory Dependencies for Human Motion Prediction Wei Mao (left) with supervisor Miaomiao Liu.