RAIL-WG : Robotic imitation learning for waypoint generation in agricultural autonomous driving

• Generalizable AI Framework: RAIL-WG introduces an imitation learning-based waypoint generation model that extends beyond agriculture and can be applied to diverse autonomous driving platforms. • Curvature-adaptive optimization: The method adjusts waypoint density by curvature, enhancing accuracy and efficiency while avoiding redundancy. • Validated across domains: Extensive simulations and real-world field tests confirm its robustness, establishing RAIL-WG as a versatile foundation for trajectory optimization in both agricultural and non-agricultural environments. Waypoint generation is a critical component of autonomous navigation, directly affecting trajectory accuracy, operational efficiency, and system robustness. Traditional fixed-interval strategies are computationally simple but lack adaptability to dynamic environments, whereas reinforcement learning (RL) methods often face unstable training and limited generalization. To overcome these challenges, we introduce robotic imitation learning for waypoint generation in agricultural autonomous driving (RAIL-WG), an LSTM-based imitation learning framework trained on expert demonstrations. Using the GROW dataset, which contains large-scale, high-resolution GPS trajectories from real-world orchard operations, RAIL-WG learns curvature-adaptive waypoint placement that balances density between straight and curved paths. Extensive simulations and field experiments show that RAIL-WG consistently outperforms both fixed-interval and RL-based baselines in trajectory tracking accuracy, computational efficiency, and smoothness. Beyond agricultural applications, the proposed framework demonstrates strong potential as a generalizable AI model for waypoint optimization, applicable to diverse autonomous systems such as mobile robots, UAVs, and ground vehicles operating in unstructured environments. This versatility highlights RAIL-WG as a scalable solution for adaptive navigation across heterogeneous domains.