The driver went open source. The person who knows what to print did not.
On September 1 a research team published LightNav-0 and put the weights on Hugging Face. It is a navigation policy built on a pretrained vision-language model, trained on about two thousand simulated scenes and four thousand hours of navigation, with no task-specific heads. The paper reports state-of-the-art monocular results across ten public navigation benchmarks. The demonstration that matters is simpler than the benchmarks: the same model, unchanged, drives a humanoid and a biped from two different makers, zero-shot.
Printers, drivers, and who knows what to print
We have used one analogy on this site since July. Robot bodies are printers; other people build them and the prices are collapsing. Motor intelligence is the driver; other people train it and we integrate whichever passes our gate. The one who knows what to print, for whom, and what must never be printed is the site mind. Bodies commoditize. Drivers commoditize. The relationship with the building compounds.
A navigation driver with open weights that runs on any body is that second sentence arriving as a download. Two months ago the cheapest way to move a robot down a corridor was the body maker's closed stack or a bought navigation package with its own sensors. Now a monocular camera and an odometry feed are enough to try, and the model does not care whose body it is on. Every body we buy or rent got cheaper to run, and the argument that the mind can be lifted off the hardware got a public proof from someone who is not us.
What a free driver cannot do
It does not know which door is a ward door. It does not know that the corridor it is entering has a resident who is not to be approached, that the linen cart it is carrying was requested by a named person, or that the request was authorized. It does not know when to stop and ask a human, and it has no record to hand a surveyor when it was in the wrong room at two in the morning. Ten simulation benchmarks are not a hospital floor on a night shift. None of that is a criticism of the model. It is a description of the layer above it, which nobody open-sourced.
That layer is where accountability lives: an identity for the machine, room-level permissions enforced in code, a human on the escalation path, and an attested record of every action. It is what a regulated facility buys, and it is what a rented brain, open or closed, cannot supply on its own.
What we do with it
The same thing we do with every driver: run it on the bench behind the gate. If an open navigation model moves a wheeled base down a corridor zero-shot, the transport job gets cheaper and the sensor bill gets shorter. Before any open model reaches a floor, its origin and license are logged the way a body's certification is logged, and it never holds safety authority; safety is deterministic code beneath the mind. The more the driver commoditizes, the more the floor pays for the thing that does not.