[news] WildCross: A Large-Scale Benchmark in Natural Environments

[Peyman Mo]

Dear AI-Robotics Community,

We would like to share our newly released large-scale benchmark for Visual/Cross-Modal place recognition and metric depth estimation in natural environments: WildCross.

WildCross continues our broader effort to advance machine learning for robotics in natural, unstructured environments. It builds on our earlier benchmark releases, Wild-Places: Large-Scale Benchmark for LiDAR Place Recognition in Unstructured Natural Environments (https://csiro-robotics.github.io/Wild-Places/) and WildScenes: Benchmark for 2D and 3D Semantic Segmentation in Natural Environments (https://csiro-robotics.github.io/WildScenes/), and now takes the next step toward visual/cross-modal perception and depth estimation in the wild.

WildCross is a large-scale cross-modal benchmark designed for robotic perception in unstructured natural environment. It provides over 476K sequential RGB frames with semi-dense metric depth and surface normal frames, aligned with accurate 6DoF poses and synchronized dense LiDAR submaps. The benchmark supports Visual Place Recognition (VPR), Cross-Modal Place Recognition (CMPR), and metric depth estimation, with a particular focus on the domain shift and variability encountered in natural environments. WildCross covers 33 km of natural forest trails collected across 14 months with multiple revisits, making it well suited for both single-session and multi-session evaluations.

We are excited to announce that WildCross has been accepted at IEEE ICRA 2026. Alongside the benchmark, we provide the paper, code, dataset access, and published checkpoints to support research in multi-modal robotic perception in natural environments.

Project page: https://csiro-robotics.github.io/WildCross/
Paper (ICRA 2026): https://arxiv.org/abs/2603.01475
Dataset: https://doi.org/10.25919/5fmy-yg37
Github Code: https://github.com/csiro-robotics/WildCross
HuggingFace Checkpoints: https://huggingface.co/CSIRORobotics/WildCross


We hope WildCross will be a useful resource for researchers working on robust localization, cross-modal representation learning, metric depth estimation, and embodied AI for field robotics. We would be very happy to hear your feedback, questions, and ideas for future extensions.

Best regards,
Peyman Moghadam



Head of Embodied AI Cluster | Senior Principal Research Scientist | CSIRO
Professor (Adjunct) | School of Electrical Engineering and Robotics | QUT
Associate Professor (Adjunct) | School of Electrical Engineering and Computer Science | UQ