Trajectron++: Dynamically-Feasible Trajectory Forecasting With\n Heterogeneous Data

Abstract

Reasoning about human motion is an important prerequisite to safe and\nsocially-aware robotic navigation. As a result, multi-agent behavior prediction\nhas become a core component of modern human-robot interactive systems, such as\nself-driving cars. While there exist many methods for trajectory forecasting,\nmost do not enforce dynamic constraints and do not account for environmental\ninformation (e.g., maps). Towards this end, we present Trajectron++, a modular,\ngraph-structured recurrent model that forecasts the trajectories of a general\nnumber of diverse agents while incorporating agent dynamics and heterogeneous\ndata (e.g., semantic maps). Trajectron++ is designed to be tightly integrated\nwith robotic planning and control frameworks; for example, it can produce\npredictions that are optionally conditioned on ego-agent motion plans. We\ndemonstrate its performance on several challenging real-world trajectory\nforecasting datasets, outperforming a wide array of state-of-the-art\ndeterministic and generative methods.\n