LF-Net: A Learning-Based Frenet Planning Approach for Urban Autonomous Driving

Learning-based approaches hold great potential for autonomous urban driving motion planning. Compared to traditional rule-based methods, they offer greater flexibility in planning safe and human-like trajectories based on human driver demonstration data and diverse traffic scenarios. Frenet planning is widely applied in autonomous driving motion planning due to its simple representation of self-driving vehicle information. However, it is challenging to select proper terminal states and generate human-like trajectories. To address this issue, we propose a learning-based Frenet planning network (LF-Net) that learns a policy to sample and select the most human-like terminal states and then generate safe trajectories. The LF-Net includes 1) a Transformer-based sub-network that encodes environmental and vehicle interaction features, 2) a classification and scoring sub-network based on cross-attention mechanisms that captures the relationship between potential terminal states and environmental features to generate the optimal terminal states set, and 3) a trajectory generator based on LQR that fits a trajectory between the selected terminal state and the initial state. Experimental results on the real-world large-scale Lyft dataset demonstrate that the proposed method can plan safe and human-like driving behavior, and performs better than baseline methods. The detailed implementation would be available at https://github.com/zyu494/LF-Net.