Metric Anything Revolutionizes 3D Perception for Robots and Cameras

A new pretraining framework called Metric Anything is making waves in 3D perception. Developed by a research team, it tackles long-standing challenges in metric depth estimation—such as sensor noise and camera biases. The work builds on recent progress in 2D vision models, showing similar scaling potential for real-world applications.

The framework was trained on a massive dataset of around 20 million image-depth pairs. This data included reconstructed, captured, and rendered 3D scenes from over 10,000 different camera models. To improve adaptability, the team introduced the Sparse Metric Prompt technique, which randomly masks depth maps to separate spatial reasoning from sensor-specific quirks.

A lightweight, prompt-free version of the model now sets new benchmarks in tasks like monocular depth estimation, camera intrinsics recovery, and 3D reconstruction. It also performs well in depth completion, super-resolution, and radar-camera fusion. When used as a visual encoder, the model boosts the spatial awareness of multimodal AI systems.

Metric Anything has already found practical uses. In industrial robotics, companies like Covariant and Amazon Robotics have tested it for bin picking and palletising since 2024. Autonomous mobile robots (AMRs), such as Boston Dynamics Stretch and Locus Robotics, are piloting the tech for warehouse navigation in 2025. Meanwhile, automotive manufacturers like Siemens and BMW are deploying monocular 3D perception for quality checks and robotic assembly this year.

To encourage further innovation, the team has released Metric Anything as open-source. Future work will focus on expanding its training data to cover even more diverse scenarios.

Metric Anything represents a step forward in handling noisy, real-world 3D data. Its open-source release allows researchers and engineers to build on the framework for new applications. Early adopters in robotics, logistics, and manufacturing are already integrating the technology into their robots and camera systems.