We develop an image-based sim-to-real transfer technique for deep reinforcement learning. First, we train a teacher model to move along a near optimal path. We then use this model to teach a student model the correct actions along with randomization. The technique bridges the sim-to-real gap, improving the driving speed and robustness of the simulator-trained student model in the real world.

The sim-to-real transfer technique we proposed can bridge the sim-to-real gap in miniature autonomous car racing. The breakthrough is that we can boost the robustness of a simulator-trained model without compromising racing lap times in the real world. We published the research result in IROS 2020 (Workshop) and achieved great success in AWS DeepRacer races.

This technique bridges the sim-to-real gap and increases the feasibility of using deep reinforcement learning for real-world applications. The autonomous driving technique can be applied to racing cars in the International Federation of Model Auto Racing (IFMAR) to make the competition more attractive. Besides, it can be applied to automated guided vehicles (AGV) in the production line for the low cost and rapid deployment. Moreover, it can be applied to the rescue and exploration missions in high-risk environments such as rough terrain, to improve the efficiency and the region of rescue.

From this work, we won AWS DeepRacer Competitions as follows:
- 3rd place among 60＋ competitors (selected among 8000＋ participants) in 2019 AWS world final held in Las Vegas.
- 1st and 3rd places among competitors (selected among 10000＋ participants) in 2020 Championship Cup of AWS DeepRacer League.

線上展網址：
https://tievirtual.twtm.com.tw/iframe/c38029ee-8c48-4c04-bf25-7af96736b37b?group=23bfb1fa-dd5b-4836-81a1-4a1809b1bae5&lang=en