Local obstacle avoidance is important for mobile robot’s movement.
Several methodologies have been suggested to solve path planning problems. They include mathematical programming, cell decomposition, roadmap approach and potential fields. These techniques prove to be inefficient due to their high computational cost and inaccuracy caused by the trapping in the local minimum.
A vehicle model is a highly nonlinear and time-varying system as far as obstacle avoidance algorithm is concerned. The algorithm must be able to adapt to a variety of changing environments. However, model-based control methods face limitations due to changing environment. Different heuristic techniques such as fuzzy logic systems and cluster optimization algorithms can overcome the weakness of the above algorithm.
Kim Thae Yong, a researcher at the Faculty of Automatics, proposed an obstacle avoidance algorithm based on Deep Q-Learning Network (DQN) that combined reinforcement learning and deep learning.
The inputs of DQN are the coordinates of a mobile robot and obstacles, the moving velocity, and the distance between them. The output is the steering angle of a mobile robot.
The trained DQN model includes both path planning and path tracking techniques and it is computationally fast and applicable to complex environments because the dynamic model of a robot is not used.
Using MATLAB, he built an environment model and a model of mobile robot and performed simulations comparing the proposed algorithm with other obstacle avoidance algorithms in various situations. The simulation results showed that the mobile robot successfully avoided obstacles after 25 000 iterations.
You can find details in his paper “Obstacle Avoidance Control of Autonomous Mobile Robot Based on Deep Q-Learning Network” in “Proceedings of KUTIC-2025”.