Metric AI Lab

Metric is a Physical AI Research Lab. We are building Botex - the execution layer for Physical AI (Codex for robotics). It is powered by our proprietary foundation model for robotic failure detection — a single evaluation layer that measures progress, flags success or failure, and explains what went wrong, for any robot. Our mission is to make robots easier to control, monitor, evaluate, and continuously improve.

Robotics today has increasingly capable AI models, but reliable execution in the physical world remains challenging. Teams rely on fragmented tools for robot control, monitoring, failure detection, and evaluation. This patchwork doesn't scale, doesn't transfer easily across platforms, and makes deploying intelligent robots unnecessarily complex.

We're developing Botex around three tightly integrated technologies: a multimodal, video-first harness for interacting with and controlling robots; a proprietary 2B edge video model for real-time execution monitoring, progress tracking, failure detection and anticipation, and safety; and a research-backed benchmarking framework for robot control and evaluation. Our approach is designed for different embodiments, from robotic arms to humanoids.

One Harness, Full Execution Loop

Instead of treating robot control, monitoring, and evaluation as separate problems, Botex connects them in a single execution loop. Robots can be controlled through text, voice, video, clicks, and direct controls, while our video model continuously observes their behavior, monitors progress, and identifies potential failures.

Our technology is designed for real-world integration — fast enough to run alongside live robot operation, adaptable across robotic platforms, and deployable in both cloud and edge environments. Built-in robot skills and execution intelligence help AI models perform physical tasks more reliably and efficiently.

Built for Real-World Impact

We're building for robotics developers, VLA researchers, robot manufacturers, and integrators. Botex supports the full development lifecycle: benchmarking VLMs and VLAs, evaluating policy rollouts, controlling physical robots, and monitoring task execution in real time.

The same infrastructure can support everything from research and data collection to production robotics, where anticipating failures, monitoring safety, and understanding task outcomes become mission-critical.

Science Meets Engineering

Research excellence drives everything we do. We're building at the frontier of video understanding and embodied reasoning, combining large-scale multimodal learning with the robustness and latency required on real hardware.

Our approach is empirical and iterative, supported by research papers, real-world and simulated robotics datasets, and hands-on hardware experimentation. We measure success not only by benchmarks, but by how effectively our technology improves robotic task execution.