IEEE RA-L | UTracker: Underwater Active Target Tracking Using Egocentric RGB Images, with Motion Capture for Sim-to-Real AUV Validation

Professor Liang Hu’s team at the International Institute of Artificial Intelligence, Harbin Institute of Technology, Shenzhen, published a paper in IEEE Robotics and Automation Letters (RA-L) entitled “UTracker: Learning Visuomotor Policies for Underwater Active Target Tracking via Imitation Learning and Diffusion Model.”

The paper presents UTracker, a visuomotor policy learning framework for underwater active target tracking that integrates reinforcement learning, imitation learning, and diffusion models. A state-based expert policy is first trained in simulated underwater environments and used to generate paired image–action expert demonstrations. A visuomotor policy is then distilled from these demonstrations, relying only on egocentric RGB images to generate safe and smooth tracking actions for an AUV and enable active tracking of non-cooperative targets.

In the real-world tank experiments, the NOKOV underwater motion capture system provides real-time pose data for the AUV. The motion-capture poses are used to transform the local waypoints generated by UTracker from egocentric RGB observations into global coordinates. These are then converted via MAVROS into control commands and sent to the AUV autopilot for execution, supporting closed-loop deployment of the simulation-trained visuomotor policy on a real underwater robot and its Sim-to-Real validation.