Introduction
Deep learning and reinforcement learning are bringing changes into robotics. In IARC challenge, we designed a hierarchical decision and control for a continuous multi-target problem: policy evaluation with action delay. We further proposed Time-in-Action reinforcement learning, mapping the state-action pair to the time accomplishing the action by its underlying controller. We also used fusing raw fisheye image and attitude data to detect and locate objects in a 3D environment with a deep neural network.
Recently we utilized the deep learning method to leverage the performance of robotic navigation. We used the self-supervised learning method to complete obscured parts of obstacles and combine this map prediction during planning to accomplish fast and safe navigation.
We are looking forward to learning biased sampling in trajectory planning. We are also interested in model-based reinforcement learning, as well as learning-based low-level controllers and learning-based state estimation.
Video