Professor Sung-Eui Yoon from the school of computing at KAIST has developed a robust robot navigation method under sensor occlusions with deep reinforcement learning collaborated with assistant professor Daehyung Park in the same department.
The results of this research, led by graduated Master's student Hyeongyeol Ryu, has been published at ICRA 2022 (IEEE International Conference on Robotics and Automation), the most prestigious academic conference in the field of robotics, which was held in Philadelphia, USA from May 24 to 26 2022. (Title of the paper: Confidence-Based Robot Navigation Under Sensor Occlusion with Deep Reinforcement Learning)
As the robot comes out of the well-controlled environments such as laboratory, robot navigation using LIDAR (Light Detection and Ranging) often suffers from the unexpected occlusions on the sensor surface due to dust, water, or smudge. Such occlusions lower the visibility of the sensor and might cause potential collisions. Therefore, the research team has focused on building a navigation policy robust to the various types of occlusions.

Figure 1. An example of the sensor occlusion during robot navigation. The unexpected smudge occludes the 2D LiDAR sensor, and corresponding observation data gives empty (i.e., zero) values.

To complement the partial visibility due to the occlusions, the research team continuously updates the confidence map representing which region is recently sensed and introduces a confidence prediction network to predict which action widens the sensed area, increasing the confidence value. In addition, by integrating this information into the reinforcement learning objective and reward functions, the research team induced the policy, maximizing the confidence value while reaching the goal without collisions.
The research team expects that this research will be helpful in situations where many external variables cannot be controlled, such as in agricultural environments and indoor service robots.

Figure 2. Demonstration of our method under the sensor occlusions.

This research paper was selected as the Outstanding Navigation Award Finalist in recognition of the excellence of the research. Among the 1500 papers published at the conference, only 40 papers were selected.