Presentation Information
[A-10-27]Development of an SB3-Based Educational Environment for EV3 Line Tracing
〇Hikaru Ohkubo1, Tohgoroh Matsui1 (1. Chubu Univ.)
Keywords:
Reinforcement Learning,Deep Reinforcement Learning,Education
In this study, we developed an environment framework that enables learning and evaluation using Stable Baselines3 (SB3) for a line-tracing task on an EV3 simulator. To adapt the environment to SB3 while preserving the semantic specifications of the original environment, the environment itself generated only observations, rewards, and termination conditions, while SB3-dependent input/output transformations were separated as wrappers. Observations were obtained by normalizing front-view images from the EV3 simulator to 64×48 pixels. Actions were mapped to four steering values. Rewards were defined stepwise based on the distance between the centroid of the black line region and the image center. Using DQN implemented in SB3, ten 4,000-step training trials were conducted, and all trials produced a policy capable of line tracing. These results confirm that the proposed environment enables stable learning and evaluation of EV3 line tracing using SB3. Future work includes applying the framework to a physical SPIKE Prime robot.