We aim to go beyond the limitation of modular-based autonomous driving system which operates well only within predictable bounds that all traffic situation is defined directly by humans.
Deep neural network(DNN)-based End-to-End autonomous driving is the key to this limitation, which will be the start point of Level 5 autonomous driving.
DNN-based End-to-End driving implements autonomous driving systems as single deep neural networks without modularizing each part of the system.
It is expected for DNN-based End-to-End autonomous driving to result in better performance than modular-based autonomous driving
as it finds an optimized behavior between inputs by training abundant datasets.
Architecture of Reinforcement learning-based End-to-End autonomous driving
Reinforcement learning(RL)-based End-to-End autonomous driving can identify features that are not defined by humans, which the conventional modular autonomous driving cannot do, and develop driving solutions that cope with various undefined driving situations on its own. It will overcome the limited response capabilities of modularized autonomous driving system.
Architecture of RL-SESR
Simulation to Real (Sim2Real) technology (2021)
Obtaining training data in real-world road environments for developing optimal driving policies of reinforcement learning is very difficult because it is accompanied by a very high probability of accidents.
Owing to the limitations of gathering real-world data, simulation environments are utilized for training the agents. The distribution shift between the simulation and real-world environment(Sim-to-Real gap) causes low interpretability of the observation, degrading the performance of the policy in the real-world.
As Sim-to-Real gap has drawn considerable attention to reinforcement learning field, we are researching the Simulation to Real(Sim2Real) technology that accomplishes efficient policy transfer from simulation to the real-world.
Development of Simulation to Real(Sim2Real) technology that improves interpretability and allows driving policy obtained from simulation to perform well even in real-world road environments, which aids in providing a potentially infinite data source and alleviates high risk to get data in real-world environments.
Segmented Encoding for Sim2Real of RL-based End-to-End Autonomous Driving, IEEE Intelligent Vehicles Symposium (IV), 2022
Reinforcement Learning based End-To-End self-driving Using Variational AutoEncoder, 2020 IPNT conference , November 2020
Architecture of Imitation
Learning-based End-to-End Autonomous Driving
Imitation Learning-based End-to-end
Autonomous Driving with High Definition Map
Imitation learning(IL)-based end-to-end autonomous driving is a technology in which neural networks are trained to imitate the human driver’s driving polices, referring to observation acquired by a human driver and actions determined at that time.
Human drivers can take different actions against the same state. So, if the IL E2E network is trained by the human driver’s driving polices with this ‘decision diversity’, the IL E2E network can’t output consistent action for any state.
The problem of decision diversity is the critical limitation of IL, and we are researching to solve it.
Utilizing various traffic indicators, observations measured from multiple sensors, and high definition(HD) map, driving policies are trained not only to solve the decision diversity but also to determine almost optimal action even in complex driving scenarios.
Efficient Autonomous Driving using S-BEV image, 3rd Korea Artificial Intelligence Conference 2022
Imitation Learning-based End-to-end Autonomous Driving with High Definition Map, 2nd Korea Artificial Intelligence Conference, 2021