LAST UPDATED · JULY 27, 2026FOUNDER-LED · PRE-MODEL · NO DEPLOYED VERTEX SYSTEM
Company fact sheet
Location
Fully remote — United States
Founder & CEO
Romir Patel
Mission
A target generalist policy for robots in extreme-purpose contexts—fireground, EMS scene, search-and-rescue, and defense logistics—with hazmat as a planned future domain. One backbone, transferable across hardware, supervised by a human operator.
Stage
Pre-model research. Building infrastructure, collecting data, standing up the training stack. No deployed model yet.
Customers
Initial groups: Firefighters · Combat medics · Search-and-rescue crews · Defense logistics operators. Planned future domain: Hazmat.
Published work
Mesquite MoCap (arXiv:2512.22690) — affordable, browser-deployable real-time motion capture.
Open work
Colosseum — a
two-axis streaming-telemetry benchmark for doctrine-grounded AI. The v0.0.4 paper describes
46 rules,
37 hand-authored scenarios,
1,200+ procedural variants,
three executable domains, and 226 tests. It reports 149 open-book trajectories, a 90-episode closed-book audit, and a separate 27-episode overload study. Matching public artifact sync is pending. Single-vendor, small-n, 0/46 SME sign-offs:
research instrument, not deployment certification.
Approach
Train one foundation model on large-scale sensorimotor data (teleop, motion capture, simulation rollouts). Apply the recipes that built foundation models in language and vision to produce policies that generalize across embodiments. Operate from first principles — we don’t copy architecture choices because they worked somewhere else.
Safety posture
Before any deployment, Vertex requires an independent safety layer, a real-time human-override path the policy cannot disable, staged real-world validation, and independent review at each gate. These are pre-deployment requirements, not claims of a currently fielded implementation.