The same vision-language-action loop runs in SimArena and on a real ROS 2 stack. How the runtime keeps the two honest.
The promise is simple: the policy you validated in SimArena is the artifact that ships to your robot. No re-implementation, no porting layer that quietly changes behavior. One binary, sim to silicon.
Keeping that honest is mostly about discipline at the boundary. The runtime exposes the same observation and action interface whether it's driving a simulated embodiment or a ROS 2 stack, and it refuses to run a policy whose I/O signature doesn't match the target.
We target the open VLA lineage — RT-2-X, OpenVLA, π0, NORA — plus diffusion-policy and ACT baselines, so you're not locked into one model family. Bring your own checkpoint.
The result is that 'works in sim' and 'works on hardware' are finally the same sentence about the same file.