onboard autonomy lives in the vehicle. offboard lives in the site: cameras, compute, networking built into the place, driving devices that don't carry the brains themselves.
why it hasn't happened yet
- who pays. onboard puts the cost in the product, so the buyer pays per unit. offboard needs a site willing to install and maintain cameras, compute, networking. stuttgart works because a garage operator and two big companies, including bosch, built it together. someone has to want that bill.
- has to work everywhere. a car or robot sold to the public can't assume the right infrastructure exists wherever it goes. onboard works anywhere, so that's where the investment went.
- no shared interface. for a site to drive anyone's device, the device needs a standard control interface and the site needs to know its dynamics. the only standards i found are for narrow cases, like vehicle marshalling.
- reliability and safety. offboard control depends on the network. drop the connection and the device still has to stop itself safely, so it needs some onboard safety logic anyway. that limits how "dumb" the device can actually be.
- latency. control loops that need a fast reaction are easier to run on the device than across a network.
- accountability. if the site's system causes a crash, who's liable isn't clear yet.
none of that kills the idea, it explains why onboard won first. a closed site, one operator, enough devices to share the cost across, that's the setup where offboard wins.