ABOUT WORKSHOP

AI-based perception for autonomous driving is an important topic in both academia and industry as it sits at the boundary between state-of-the-art learning methods and safety-critical real-world deployment. This workshop aims to bring together work on making perception more reliable under real conditions including robustness towards anomalies, out-ofdistribution (OOD) situations, stronger 3D perception including bird’s eye view (BEV) projection and depth estimation, and evaluation practices that better reflect safety and deployment needs. We also welcome work on multimodal perception and sensor fusion (e.g., camera/LiDAR/radar), and recent directions such as foundation models and vision-language models for autonomous driving perception. The theme fits well within IJCAI/ECAI’s scope by connecting core AI challenges (generalization, uncertainty, robustness, scalable learning, and evaluation) with a high impact application domain where these challenges are clearly visible. Autonomous driving is also a context where perception models can look strong on benchmark test sets but still show reliability problems after deployment. Performance can drop with changes in location, weather, lighting, traffic patterns, sensor noise, partial sensor failure, or calibration and rig changes. These issues are strongly related to safety, because failure often comes from distribution shifts, long-tail events, and poorly captured corner cases rather than average conditions. As a result, progress in this area needs not only new models, but also clearer problem definitions, better test approaches, and more informative metrics. We invite contributions across a broad range of perception tasks and settings, including robust detection, tracking, segmentation, occupancy and mapping, OOD/anomaly detection and openworld perception, 3D perception and BEV-centric methods, depth estimation and depth completion for driving, and learning methods that improve robustness such as self-supervision, test-time adaptation, and data-centric pipelines (including synthetic data and targeted long-tail collection). We also encourage work on rig invariance, sensor fusion, calibration and drift, and practical deployment monitoring, because these are common sources of real-world performance drops that are often underexplored in academic evaluations. Overall, the workshop aims to connect these research directions into a shared discussion on how to build and evaluate perception systems for real driving conditions, aligning with IJCAI/ECAI’s focus on core AI questions.