Technical Name Deep Reinforcement Learning with Action SmoothnessIts Application to Autonomous Miniature Car Racing
Project Operator National Yang Ming Chiao Tung University
Project Host 吳毅成
We propose a sim-to-real transfer technique utilizing a generative adversarial model CycleGAN for virtual-to-real image style transformation, reducing the gap between pure visual self-driving model performance in simulationreality. Then, we propose a method for action smoothness by enhancing the continuity in consecutive actions. The combination of these methods successfully improves the speedstability of the physical race car.
Scientific Breakthrough
We propose a sim-to-real transfer techniquean action smoothing method which greatly enhances the driving performance of DRL-based self-driving car racing by addressing challenges related to virtual-to-real environment differencescontrol stability.This technology was presented at workshops of two toptier conferences, ICRAIJCAI, in 2022,also achieved a remarkable accomplishment by winning the top three in the AWS DeepRacer League competition (with 150,000+ contestants).
Industrial Applicability
We propose a sim-to-real transfer techniquean action smoothing method that enhances the feasibility of DRL in real-world applications, opening up possibilities for robotics, robotic arms,unmanned drone control in real-world settings. Furthermore, this technology finds applications beyond self-driving car competitions, including unmanned transport vehicles in factoriesemergency exploration in challenging environments. It offers cost-effectivefast deployment capabilities.
Keyword Deep Reinforcement Learning Deep Learning Network Autonomous Car Racing Sim-To-Real Transfer End-to-end learning Soft-Actor-Critic (SAC) Action Policy Smoothness CycleGAN
  • Contact
  • Cindy Ko
other people also saw