[Paper Review] GAMMA: A General Agent Motion Model for Autonomous Driving

Yuanfu Luo, Panpan Cai+2 more|arXiv (Cornell University)|Jun 4, 2019

Traffic Prediction and Management Techniques4 citations

TL;DR

GAMMA is a real-time, analytical motion prediction model for autonomous driving that formulates agent motion as constrained optimization in velocity space, incorporating kinematic, collision, and environmental constraints while conditioning on human behavioral states. It achieves sub-millisecond inference speed and high accuracy on heterogeneous traffic benchmarks, enabling scalable simulation and real-time planning under uncertainty.

ABSTRACT

This paper presents GAMMA, a general motion prediction model that enables large-scale real-time simulation and planning for autonomous driving. GAMMA models heterogeneous, interactive traffic agents. They operate under diverse road conditions, with various geometric and kinematic constraints. GAMMA treats the prediction task as constrained optimization in traffic agents' velocity space. The objective is to optimize an agent's driving performance, while obeying all the constraints resulting from the agent's kinematics, collision avoidance with other agents, and the environmental context. Further, GAMMA explicitly conditions the prediction on human behavioral states as parameters of the optimization model, in order to account for versatile human behaviors. We evaluated GAMMA on a set of real-world benchmark datasets. The results show that GAMMA achieves high prediction accuracy on both homogeneous and heterogeneous traffic datasets, with sub-millisecond execution time. Further, the computational efficiency and the flexibility of GAMMA enable (i) simulation of mixed urban traffic at many locations worldwide and (ii) planning for autonomous driving in dense traffic with uncertain driver behaviors, both in real-time. The open-source code of GAMMA is available online.

Motivation & Objective

• Address the challenge of real-time, high-fidelity motion prediction for heterogeneous, interactive traffic agents in complex urban environments.
• Overcome limitations of data-driven models requiring large-scale training and analytical models lacking behavioral flexibility.
• Enable scalable, real-time simulation and planning for autonomous vehicles by combining computational efficiency with accurate uncertainty-aware predictions.
• Model diverse human driving behaviors—such as attention, intent, and aggression—explicitly through optimization parameters.
• Support downstream applications like crowd-driving simulation and belief-space planning under uncertainty using efficient, differentiable motion modeling.

Proposed method

• Formulate motion prediction as a constrained optimization problem in the velocity space of each agent, maximizing driving performance under kinematic, collision, and environmental constraints.
• Model agent geometry as polygons and use reciprocally-optimal collision avoidance to define linear constraints, enabling efficient solution via quadratic programming.
• Condition the optimization on human behavioral states (e.g., intended route, attention) as parameters, allowing flexible simulation of diverse driving behaviors.
• Use Bayesian inference to infer posterior distributions over behavioral states from observed trajectories, enabling distributional motion predictions with uncertainty quantification.
• Integrate GAMMA into a high-fidelity simulator (SUMMIT) for large-scale urban traffic simulation across OpenStreetMap-based maps.
• Employ GAMMA as a forward model in a POMDP-based planner for autonomous vehicles, using belief tree search with Monte Carlo simulations to handle uncertainty in agent intentions.

Experimental results

Research questions

• RQ1 Can a training-free, analytical motion model achieve high prediction accuracy across diverse, heterogeneous traffic scenarios while maintaining real-time performance?
• RQ2 How effectively can behavioral state parameters (e.g., intent, attention) be integrated into a constrained optimization framework to model flexible human driving behaviors?
• RQ3 To what extent does GAMMA’s computational efficiency enable large-scale simulation and real-time planning in complex urban environments?
• RQ4 How does GAMMA’s uncertainty-aware prediction improve performance in belief-space planning for autonomous vehicles in dense, unpredictable traffic?
• RQ5 Can GAMMA outperform state-of-the-art data-driven and analytical models in both accuracy and inference speed on standard benchmarks?

Key findings

• GAMMA achieves sub-millisecond inference time per agent, enabling real-time simulation of mixed urban traffic at scale across diverse global urban maps.
• On real-world benchmark datasets, GAMMA outperforms multiple state-of-the-art analytical and data-driven models in both homogeneous and heterogeneous traffic scenarios.
• The integration of GAMMA into the SUMMIT simulator enables high-fidelity, large-scale simulation of traffic agents on any OpenStreetMap-based urban map.
• In belief-space planning, POMDP-GAMMA achieves the lowest collision rate, fastest time to goal, and smoothest driving behavior compared to Rollout-GAMMA and POMDP-Simple.
• The use of Bayesian inference with GAMMA allows for robust uncertainty quantification in motion predictions, which is critical for safe and efficient autonomous planning.
• Despite simplifying assumptions (static, independent behavioral states), GAMMA maintains strong performance, with potential for future improvement via joint inference-optimization or dynamic state modeling.