HRDX: A Large-Scale Vector HD-Map Dataset

A research team has introduced HRDX, a large-scale dataset for constructing vectorized high-definition maps, spanning about 40 hours of driving data across 1,400 km [1][2]. The dataset pairs multi-sensor ground recordings with aligned aerial imagery to support autonomous-driving research [2]. The dataset captures minimally overlapping drives using six synchronized surround cameras, a 128-beam LiDAR, and centimeter-level RTK GNSS/IMU positioning [1][2]. It is further complemented by precisely aligned aerial orthoimagery, a modality absent from most prior public HD map collections [2]. Annotations cover 10 vector map classes and more than 20 semantic and topological attributes, creating a richer ontology than earlier benchmarks [1][2]. To evaluate this expanded label set, the authors introduce the Composite Score, a metric that jointly assesses geometric fidelity and attribute correctness [1][2]. Benchmark experiments indicate that HRDX's scale improves online vector-map construction, and that aligned aerial imagery provides a useful structural prior [2]. Using aerial imagery at training or inference time lifts geometric map quality, while aerial-augmented teacher models can transfer part of that benefit to camera-only student models without increasing inference-time sensor requirements [2]. The dataset and benchmarks are publicly available through the Honda Research Institute's GitHub repository [1][2]. HRDX arrives as the autonomous-driving community grapples with the limits of existing HD map resources. Prior public datasets have been constrained in geographic coverage, semantic depth, and sensor variety, often omitting overhead imagery that could inform bird's-eye-view perception [2]. By releasing a corpus several times larger than previous offerings, the team aims to support reproducible research on large-scale HD-map learning, multimodal BEV fusion, and training-time privileged information [2]. The work was submitted to arXiv on 11 June 2026 by Sahith Reddy Chada and collaborators [1]. While the primary paper does not include direct quotations, the abstract states that "reliable autonomous driving requires vectorized HD maps that are geometrically accurate, semantically rich, and scalable to long-horizon driving" [2]. The release adds to a growing body of open-source driving datasets that seek to standardize evaluation and accelerate progress in self-driving perception [2].