methods and boosting development efficiency by 5x.
• Designed and implemented a Cumulative Decoder, reconstructing the Decoder Head for multimodal trajectories,
resulting in a 48% improvement in recall and a 33% increase in mADE50 on target scenario datasets.
• Created and integrated Landmark-Loss functions to optimize the end-to-end trajectory prediction model, resulting in a
significant performance boost, with a 28% improvement in precision and a 36% increase in recall.
• Addressed real-world challenges such as trajectory compliance and centering in road tests by employing data mining
techniques and optimizing key features, completing self-closed-loop verification and deployment.