No model
yet.
We’re in early-stage research. This page is where the model card will live once we have one. For now, it’s a plain accounting of where we are and where we’re going.
Vertex is pre-model. We do not have a deployed foundation model, and we do not publish benchmark numbers we have not actually measured. We’re building infrastructure, collecting data, and standing up the training stack.
When we’ve trained something worth talking about, this page will host the full model card — architecture, training data, evaluation, capabilities, and known limitations. Until then, we’re going to say less, not more.
A single generalist policy for robots operating in extreme-purpose contexts — fireground, EMS scene, search-and-rescue, hazardous-materials response, defense logistics. One backbone, transferable across hardware platforms, supervised by a human operator at all times.
The hard parts: cross-embodiment transfer, real-world data efficiency, closing the sim-to-real gap for contact-rich tasks, and producing safety guarantees that hold up in chaotic environments rather than only in a lab.
Colosseum v0.0.4 working paper, 46-rule site index, and published methodology. Matching public artifacts and SME validation remain incomplete.
Sensorimotor data collection, simulation, and training-stack work. No Vertex foundation model has been announced.
Publish architecture, data accounting, capability tests, and failure analysis only after a real model exists and the measurements can be reproduced.
Independent safety review, hardware override, controlled trials, and domain-expert sign-off. No calendar date is promised.
We will publish a complete model card the day we deploy. We will not invent benchmark scores. We will document failure modes alongside capabilities. We will not use this page to advertise — only to report.