CRAT-Pred: Vehicle Trajectory Prediction with Crystal Graph Convolutional Neural Networks and Multi-Head Self-Attention

Julian Schmidt, Julian Jordan, Franz Gritschneder, Klaus Dietmayer

Abstract

Predicting the motion of surrounding vehicles is essential for autonomous vehicles, as it governs their own motion plan. Current state-of-the-art vehicle prediction models heavily rely on map information. In reality, however, this information is not always available. We therefore propose CRAT-Pred, a multi-modal and non-rasterization-based trajectory prediction model, specifically designed to effectively model social interactions between vehicles, without relying on map information. CRAT-Pred applies a graph convolution method originating from the field of material science to vehicle prediction, allowing to efficiently leverage edge features, and combines it with multi-head self-attention. Compared to other map-free approaches, the model achieves state-of-the-art performance with a significantly lower number of model parameters. In addition to that, we quantitatively show that the self-attention mechanism is able to learn social interactions between vehicles, with the weights representing a measurable interaction score. The source code is publicly available.

1. Introduction

Accurate vehicle trajectory prediction is a cornerstone of safe autonomous driving. While many state-of-the-art (SOTA) models rely heavily on high-definition (HD) maps to understand road geometry, such information is not always reliable or available due to localization failures or outdated data. CRAT-Pred (Crystal Attention Prediction) addresses this by providing a robust "map-free" prediction framework. By leveraging spatial relationships and social interactions through a novel architecture, CRAT-Pred achieves competitive performance with a fraction of the parameters required by traditional map-heavy models.

2. Methodology

CRAT-Pred utilizes a three-stage pipeline to process vehicle data and generate multi-modal future trajectories.

2.1 Input Encoder

The model first processes the past two seconds of a vehicle's motion. A Long Short-Term Memory (LSTM) network encodes these discrete displacements into a high-dimensional hidden state, capturing the temporal dynamics of each actor in the scene.

2.2 Interaction Module

The core of CRAT-Pred is its interaction module, which models social dependencies between vehicles using two primary mechanisms:

  • Crystal Graph Convolutional Neural Networks (CGCNN): Inspired by material science models of atomic bonds, CGCNNs are used to model vehicle interactions. Unlike standard graph convolutions, CGCNNs explicitly incorporate edge features—specifically the relative Euclidean distance between vehicles—to update node representations.
  • Multi-Head Self-Attention (MHSA): This layer allows the model to dynamically weigh the importance of surrounding vehicles. It enables the system to "pay attention" to critical actors (e.g., a vehicle merging ahead) regardless of their absolute distance.
Figure 1: Overview of the CRAT-Pred architecture, illustrating the flow from LSTM encoding through the CGCNN/Attention interaction module to the multi-modal decoder.

2.3 Output Decoder

The final stage employs linear residual layers to generate multiple possible future trajectories. Training follows a two-stage process: an initial single-path backbone training followed by the addition of multiple decoders trained with a "Winner-Takes-All" loss to ensure diverse and realistic predictions.

3. Experimental Results

The model was evaluated on the Argoverse Motion Forecasting Dataset, a benchmark containing over 300,000 real-world driving sequences. Despite the availability of HD maps in the dataset, CRAT-Pred intentionally ignores them to demonstrate its map-free efficacy.

3.1 Quantitative Performance

CRAT-Pred achieved a state-of-the-art minADE (minimum Average Displacement Error) of 1.82m on the Argoverse test set. As shown in Figure 2, this performance significantly outperforms standard map-free baselines and approaches the accuracy of several map-inclusive models.

Figure 2: Performance comparison on the Argoverse benchmark. CRAT-Pred (Ours) demonstrates superior accuracy compared to other map-free models and remains competitive with map-based approaches.

3.2 Model Efficiency

A key advantage of CRAT-Pred is its architectural efficiency. The model operates with approximately 515k parameters. In comparison, competing models like LaneGCN and WIMP utilize over 1 million and 20 million parameters, respectively. This efficiency makes CRAT-Pred highly suitable for real-time deployment on embedded automotive hardware.

4. Discussion and Conclusion

CRAT-Pred demonstrates that sophisticated social interaction modeling can compensate for the absence of HD maps. The quantitative analysis of self-attention weights reveals that they serve as a superior proxy for social interaction compared to simple distance metrics. While the model currently lacks geometric constraints (which can lead to physically possible but traffic-illegal trajectories), it provides a powerful foundation for robust, map-independent navigation. Future work may explore hybrid architectures that integrate map features into the CGCNN framework to further enhance prediction accuracy in complex urban environments.

Summary updated

Abstract

Predicting the motion of surrounding vehicles is essential for autonomous vehicles, as it governs their own motion plan. Current state-of-the-art vehicle prediction models heavily rely on map information. In reality, however, this information is not always available. We therefore propose CRAT-Pred, a multi-modal and non-rasterization-based trajectory prediction model, specifically designed to effectively model social interactions between vehicles, without relying on map information. CRAT-Pred applies a graph convolution method originating from the field of material science to vehicle prediction, allowing to efficiently leverage edge features, and combines it with multi-head self-attention. Compared to other map-free approaches, the model achieves state-of-the-art performance with a significantly lower number of model parameters. In addition to that, we quantitatively show that the self-attention mechanism is able to learn social interactions between vehicles, with the weights representing a measurable interaction score. The source code is publicly available.
BibTeX
@inproceedings{schmidt2022crat,
  author = {Julian Schmidt and Julian Jordan and Franz Gritschneder and Klaus Dietmayer},
  title = {{CRAT-Pred: Vehicle Trajectory Prediction with Crystal Graph Convolutional Neural Networks and Multi-Head Self-Attention}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2022},
  url = {https://openreview.net/forum?id=tizBYtoWgp},
}
Figure