Behavior Encoder Transformer BETR: A Transformer Encoder for Predicting Agent Behavior
Mahmoud a Elhusseni, Yara Maher, Emad El‐Sayed, Mohamed F. Abdelkader · Port-Said Engineering Research Journal · 2025
Anticipating the future behavior of road users stands as one of the most formidable challenges in the realm of autonomous driving. Achieving a comprehensive understanding of the dynamic driving environment requires an autonomous vehicle to accurately predict the motion of other traffic participants within the scene. As the complexity of motion prediction tasks increases, capturing intricate spatial relationships, temporal dependencies, and nuanced interactions between agents and map elements becomes crucial. Our proposed hierarchical architecture strategically incorporates transformers, effectively modeling both local and global representations to extract multi- scale features. Leveraging the potency of transformers, BETR adeptly captures and encodes intricate patterns of agent interactions, spatial dependencies, and temporal dynamics. Demonstrating superior performance in predicting agent behavior compared to conventional methods, BETR proves its efficacy through extensive experiments. Its capacity to adapt to diverse scenarios establishes BETR as a robust and versatile solution for the intricate task of agent behavior prediction.