Improving Multi-Agent Motion Prediction with Heuristic Goals and Motion Refinement

Carlos Gómez Huélamo, Marcos V. Conde, Rafael Barea, Luis Miguel BergasaWorkshops @ CVPR 2023 · DOI

Abstract

Predicting the future trajectories of surrounding agents is a fundamental challenge for autonomous vehicles. This paper introduces Argo2GoalMP, an efficient end-to-end pipeline that combines deep learning architectures—specifically Transformers and Graph Neural Networks—with physics-based heuristic scene understanding. Evaluated on the Argoverse 2 benchmark, the model achieves state-of-the-art results while significantly reducing computational overhead, offering a 41% improvement in inference speed and a 20% reduction in parameter count compared to previous leading models.

Introduction

Motion Prediction (MP) serves as the predictive engine for autonomous driving systems, enabling vehicles to anticipate the movements of cars, cyclists, and pedestrians. To plan safe paths, a model must navigate two primary hurdles: the inherent multimodality of human behavior (e.g., a driver choosing between turning or going straight) and the strict real-time processing constraints of vehicle hardware. While data-driven "black box" models have become standard, they often struggle with physical consistency or excessive computational demands. Argo2GoalMP addresses these issues by injecting rule-based heuristic knowledge into a lightweight, high-performance deep-learning framework.

Methodology: The Argo2GoalMP Pipeline

The proposed architecture utilizes a multi-stage approach to refine predictions from raw sensor and map data into smooth, plausible trajectories.

Social and Map Encoding

The system replaces traditional, memory-intensive LSTMs with a Transformer-based Social Encoder to process agent histories. Simultaneously, road geometry is handled by a LaneGCN backbone, which represents lane centerlines and their connectivity as a structured graph.

Heuristic Goal Proposals

A core innovation of this work is the use of heuristic "hints." Before the neural network generates final trajectories, the system applies kinematic filters (such as Savitzky-Golay) and performs a Depth First Search (DFS) on the map to identify physically plausible future lanes. This regularizes the search space, allowing the model to focus on high-level reasoning rather than relearning basic road rules.

Global Interaction via Crystal-GCN

To model social dynamics—such as yielding at a roundabout—the authors employ a Crystal-GCN. Unlike standard graph layers that focus primarily on node features, Crystal-GCN explicitly incorporates "edge features," such as relative distance and orientation between agents. This ensures that spatial relationships are treated as primary inputs for interaction modeling.

Motion Refinement

The final stage involves an MLP-based module that predicts a "residual" correction for the initial trajectories. This step ensures that the final output is both temporally consistent and kinematically smooth.

Key Results

The model was rigorously tested on the Argoverse 2 (AV2) dataset, which spans over 2,000 km of diverse urban driving data.

Multimodal trajectory predictions generated by the Argo2GoalMP model in a complex traffic scenario.

Argo2GoalMP achieved top-tier performance on the AV2 leaderboard, particularly in minADE (Average Displacement Error) and minFDE (Final Displacement Error) metrics. Beyond accuracy, the model demonstrates superior efficiency:

  • Parameter Efficiency: The model utilizes 4.92 million parameters, a 20% reduction from the 6.2 million used by the GANet baseline.
  • Inference Speed: Runtime was reduced by approximately 41%, facilitating deployment on standard autonomous vehicle compute platforms.

Discussion and Future Work

The success of Argo2GoalMP highlights the value of combining data-driven learning with domain-specific heuristics. By providing the model with a "hint" of road geometry, the architecture achieves higher precision with fewer parameters. Future research directions include the development of map-adaptive loss functions to further penalize lane deviations and the use of knowledge distillation to compress the model for low-power edge devices. While the current reliance on heuristic quality is a potential limitation in off-road or unmapped scenarios, the model sets a new standard for efficient, real-time motion forecasting in structured urban environments.

Summary updated

BibTeX
@inproceedings{huelamo2023improving,
  author = {Carlos Gómez Huélamo and Marcos V. Conde and Rafael Barea and Luis Miguel Bergasa},
  title = {{Improving Multi-Agent Motion Prediction with Heuristic Goals and Motion Refinement}},
  booktitle = {Workshops @ CVPR},
  year = {2023},
  doi = {10.1109/cvprw59228.2023.00561},
}
Figure