The article presents AnyD, a geographically-aware conditional imitation learning model for autonomous driving. AnyD can efficiently learn from diverse and globally distributed data, including dynamic environmental, traffic, and social characteristics. It uses a high-capacity geo-location-based channel attention mechanism to adapt to local nuances and model similarities among regions. This approach is scalable and can handle imbalanced data distributions and location-dependent events. AnyD outperforms existing models by over 14% in open-loop evaluation and 30% in closed-loop testing.
Publication date: 21 Sep 2023
Project Page: https://arxiv.org/abs/2309.12295v1
Paper: https://arxiv.org/pdf/2309.12295
