LEO: A Graph Attention-Based Framework for Learned Object Extensions and Adaptive Sensor Fusion for Autonomous Driving Applications

ICLR 2026 Conference Submission17321 Authors

Keywords: Graph Attention Network, Multi Sensor Fusion, Automated Driving, Autonomous Vehicles, Shape Estimation

Abstract: Accurate shape and trajectory estimation of dynamic objects is a fundamental requirement for reliable perception in Automated Driving (AD). In the classical versions of AD algorithms and stacks, various Bayesian extended object geometric models are used to provide object-related extensions and trajectories. Performance of such approaches are deeply connected with the completeness of a-priori and update-likelihood functions. Recent deep learning approaches improve flexibility by learning shape features directly from raw or fused sensor data, but they often rely on dense annotated datasets and high computational resources, which restricts their applicability in production vehicles. We aim to improve production-level automated driving systems by integrating the computational efficiency and theoretical robustness of geometric methods with the adaptability and generalization capabilities of modern deep learning techniques. We employ a task-specific parallelogram-based ground-truth formulation to represent object extensions, facilitating expressive modeling of complex geometries such as articulated trucks and trailers. Our primary contribution is the development of a novel spatio-temporal Graph Attention Network (GAT)-based model, Learned Extension of Objects (LEO), that demonstrates proficiency in adaptive fusion weight learning, temporal consistency, and multi-scale shape representation from multi-modal production grade sensors. LEO successfully generalizes across various sensor modalities, configurations, object classes, and geographic regions, exhibiting robustness even under challenging conditions and longer range targets. We have presented these observations and evaluations based on the real-world Mercedes-Benz SAE Level-3 (L3) DRIVE PILOT dataset in our article. Furthermore, its computational efficiency makes it a suitable candidate for integration into a real-time production system, although further validation and integration efforts are necessary for deployment in safety-critical systems.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 17321