Projects

Reinforcement learning-based navigation for indoor service robots (2022.1-present)

• The control commands of the robot is output directly by a network that encodes the features of laser, goal and speed.
• The network is trained based on reinforcement learning algorithms like PPO and Decision Transformer with simulators like Stage, Gazebo and Webots.
• Some security protection strategies and traditional path planning algorithms are combined to solve some corner cases that reinforcement learning cannot handle.
• The runtime code is written in C++ and deployed on Nvidia TX2 with ROS.

Navigation of agricultural inspection/delivery robots (2021.6-2022.9)

• The Circular Accessible Depth is deployed as the perception module to predict the traversability.
• The UGV need to travel between the warehouse and the field. We built a point cloud map for indoor localization, which uses fast_lio_localization, an open-source project that I maintained.
• However, due to the influence of tree growth on point cloud maps, we turn to fusing the odometry, IMU and GPS by Kalman filter for localization in the field.
• The navigation framework was deployed on Nvidia TX2 with ROS.

Object detection-based people counting with edge computing (2020.9-2021.5)

• YOLOv5 is taken as the detection network and trained to detect human body or head according to the characteristics of different application scenarios.
• The trajectories of bodies or heads are computed by some tracking algorithms like SORT.
• Then, the number of people who pass the line set by user could be counted by calculating the geometric relationship between the trajectories and the line.
• The YOLOv5 model accelerated by TensorRT and the tracking code were deployed on TX2 and can run at more than 10 frames per second.