GAMMA: A General Agent Motion Prediction Model for Autonomous Driving

Abstract

Autonomous driving in mixed traffic requires reliably predicting the motion of nearby traffic agents, such as pedestrians, bicycles, cars, buses, etc. . This prediction task is extremely challenging, because of the diverse geometry and dynamics of traffic agents, multi-way interactions among them, and complex road conditions. This paper presents GAMMA, a general agent motion prediction model for autonomous driving. GAMMA predicts the motion of heterogeneous traffic agents with different geometric, kinematics, and behavioral constraints. It formalizes motion prediction as constrained geometric optimization in the velocity space and integrates both physical and behavioral constraints into a unified framework. Our results show that GAMMA outperforms state-of-the-art motion prediction methods substantially on real-world datasets/