Primary subject
Artificial intelligence
Models whose output is a movement, a stop, or a decision a machine has to make in time. Language models are included only when they change what a robot is instructed to do, or when they plan above a set of smaller controllers.
On the machine
Lowest delay and the tightest power budget. The model is usually smaller than the one in a research demo.
In the building
A local server shared by several machines that already trust the site. Common in factories.
From a network
Paid inference on someone else’s computer when the robot cannot carry the model. Only real if latency and failure are specified.
What to separate in an announcement
The policy (what action is chosen), the data (whose sites and failures it learned from), the computer (on the robot, in the building, or in a distant data centre), and the envelope (the tasks the operator will actually allow). Those four are often presented as one product. They are not.
Text from the public internet does not replace contact, glare, force or streets the machine has not seen. Sensor networks matter to this section when a robotics team could license the data, and when the sample can be checked.
Latest
Background
- AI · BackgroundPhysical AIA language model can describe a doorway. A robot has to pass through it without taking the frame along.
- AI · BackgroundWhere inference runsSome decisions cannot wait for a data centre. The model has to live on the machine, or on a computer close to it.
- AI · BackgroundTraining data from the physical worldText on the internet is not a substitute for contact, glare, audio and streets the robot has not seen.