在地平线搞自动驾驶的这三年

Core Insights - The article discusses the transition from autonomous driving to embodied intelligence, highlighting the differences in challenges and solutions between the two fields [2] - It emphasizes the importance of documenting past experiences in autonomous driving, even if they did not receive widespread attention, as they may provide practical insights for others in the field [2] Research Areas Summary - Sparse4D Series: A multi-sensor fusion perception framework that challenges the conventional BEV (Bird's Eye View) approach, arguing that it does not significantly enhance information while incurring high computational costs. The Sparse4D series aims to achieve efficient perception through sparse queries and projections [6][7] - SparseDrive: An attempt to extend the capabilities of the Sparse4D model into end-to-end planning, integrating online mapping and motion planning tasks. It successfully executed five tasks, including detection and tracking, but faced challenges in closed-loop performance evaluation [13][15] - EDA & UniMM: EDA introduces a dynamic anchor strategy for trajectory prediction, improving model convergence and accuracy. UniMM unifies existing traffic flow simulation models, addressing key performance factors in agent simulation [16][20] - DriveCamSim: A sensor simulation system designed to evaluate autonomous driving models efficiently. It focuses on generating sensor signals with high fidelity and controllability, addressing the limitations of traditional physical engine-based simulations [22][24] - LATR: A foundational model for intelligent driving that leverages large datasets for unsupervised training, aiming to understand the semantics of driving scenarios. It integrates multiple tasks into a unified framework, demonstrating effective performance across various driving tasks [26][27] Conclusion and Future Outlook - The seven modules discussed form the core link of the autonomous driving system, indicating a correct technological path. The industry is moving towards maturity in end-to-end models, with significant performance improvements for companies adopting these approaches. Future developments should focus on efficient evaluation systems and the potential of reinforcement learning to enhance model performance [30][31]