Delving into Mapping Uncertainty for Mapless Trajectory Prediction

Status: Accepted to IROS 2025 🎉
Role: Team Leader

Abstract

We investigated the effect of online-generated High-Definition (HD) map uncertainty on mapless trajectory prediction in autonomous driving, identifying the vehicle’s kinematic state as a key overlooked factor. Our approach achieved up to 23.6% performance gain over prior SOTA methods.

Key Contributions

• Novel Uncertainty Model: Proposed lightweight, self-supervised Proprioceptive Scenario Gating module that adaptively integrates map uncertainty into trajectory prediction based on ego vehicle’s future motion dynamics
• Advanced Geometric Modeling: Designed Covariance-Based Map Uncertainty model using 2D Gaussian distributions to better capture road geometry and improve robustness over prior Laplace-based approaches
• Comprehensive Benchmarking: Reproduced and benchmarked four SOTA online map construction models (MapTR, MapTRv2, MapTRv2-Centerline, StreamMapNet) and integrated with two representative trajectory predictors (Transformer-based HiVT, GNN-based DenseTNT)
• Significant Performance Gains: Achieved up to 23.6% performance improvement over prior SOTA methods in trajectory prediction metrics (minADE, minFDE, Miss Rate) on nuScenes dataset
• Efficiency Advantages: Ablation studies showed proprioceptive gating outperforms exteroceptive CLIP/ResNet-based alternatives with 10-30x faster inference speed, enabling real-time deployment of mapless trajectory prediction systems

Technical Innovations

• Identified vehicle’s kinematic state as critical overlooked factor in map uncertainty
• Full-stack evaluation framework combining map construction and trajectory prediction
• Real-time deployment optimization for autonomous driving applications