From February 2017 to September 2017, I participated in undergraduate internship program at Naver Labs Robotics group. During the 6 months of intership program, our team developed a small-sized, stand-alone and autonomous mobile robot system for indoor object collection.

Our team developed three versions of autonomous mobile robot.
(left: first prototype of mobile robot, center: second version mobile robot TT-bot, right: final version mobile robot TT2-bot)


For object collection task, I composed the software system with four steps. To improve object recognition performance deep convolutional neural network are used in perception step and deep reinforcement learning are used for motion planning.

In the second verison mobile robot, TT-bot, a target object is detected with LBP-feature based cascade classifier, and the shortest path toward the target object is generated with A* algorithm. The results were presented in ICCAS 2017.



For the final version mobile robot, TT2-bot, perception and motion planning is substituted with deep learning cores; a deep convolutional network classifier for perception and a deep reinforcement learning for motion planning respectively. Also, I integrated deep convolutional network classifier in an asynchronous manner, to increase the system bandwidth. With the implemented multi-body tracker, perception, mapping, classifier, motion planners work asynchronously, and data flow among these components is managed by an abstracted map. In the case of deep reinforcement learning-based motion planner, the abstracted map is used as an input of the network and the direction of robotic movement is generated as an output. With the implemented robot, our team got a grand prize in 2017 Korea Intelligent Robot Competition, and the overall results of devised architecture were submitted in ICRA 2018.