Intelligent agents can achieve complex continuous control and decision-making skills by learning representations from raw perception to high-level control actions. However, end-to-end policy learning is challenging for autonomous navigation as it is often limited to simplistic road environments, navigation with no interactions, or testing in solely passive settings.
Image credit: Unsplash/Kimi Lee, free licence
A recent paper on arXiv.org presents an end-to-end framework for photorealistic simulation and training of autonomous agents in the presence of both static and dynamic agent interactions.
The learned policies can be directly transferred onboard a full-scale autonomous vehicle in the real world. Real-world experiments on a full-scale autonomous vehicle are conducted. The models demonstrate high performance and generalizability on complex tasks such as autonomous overtaking and avoidance of a partially observed dynamic agent.
Data-driven simulators promise high data-efficiency for driving policy learning. When used for modelling interactions, this data-efficiency becomes a bottleneck: Small underlying datasets often lack interesting and challenging edge cases for learning interactive driving. We address this challenge by proposing a simulation method that uses in-painted ado vehicles for learning robust driving policies. Thus, our approach can be used to learn policies that involve multi-agent interactions and allows for training via state-of-the-art policy learning methods. We evaluate the approach for learning standard interaction scenarios in driving. In extensive experiments, our work demonstrates that the resulting policies can be directly transferred to a full-scale autonomous vehicle without making use of any traditional sim-to-real transfer techniques such as domain randomization.
Research paper: Wang, T.-H., Amini, A., Schwarting, W., Gilitschenski, I., Karaman, S., and Rus, D., “Learning Interactive Driving Policies via Data-driven Simulation”, 2021. Link: https://arxiv.org/abs/2111.12137