First robot to create the "brains" of other robots

Programming is reduced 20 times.

Imagine if a robot could program another robot's brain almost on its own and do in days what would take humans years, sounds Skynet style, right? but the project by Peter Burke of the University of California turns the idea into a prototype, using generative models such as ChatGPT, Claude and Gemini to design and code a machine's control system from scratch.

The promise is brutal efficiency, the inevitable discomfort, when machines write the code that governs machines, who is in control? In Terminator, robots become self-aware and dominate the world, he says in his study, and concludes by saying, in this article we take a first step in that direction.

With three intertwined brains.

The images of Figure's humanoid robots are just to illustrate, as the researcher talks about them, the heart of the experiment is a layered pipeline that separates three brain levels, the lower brain is in the flight firmware of a drone taking care of the stabilization of the actuators, the middle brain runs a Raspberry Pi Zero 2 on board and hosts via WebGCS, a panel accessible by the network for maps, telemetry and broadcast planning in real time and the higher brain adds local autonomy.

In practice, the flow begins with a human setting the direction, requirements, hardware limitations, available APIs and the desired behavior, the AI, guided by prompts and timely corrections, decomposes tasks, writes modules, adjusts dependencies and validates the integration, there is, however, a conscious brake, human controls, complex logs and auditability and the notion that models are probabilistic and need rails.

The machine exploring the solution space

The ideal interaction here is not man against machine, but man defining objectives and limits, machine exploring the solution space and man closing the cycle with validation and approval, the same method fits in any robot that has a programmable brain, humanoids, autonomous ground vehicles in industrial yards, arms that learn repetitive tasks on the line, pipeline inspection ships, even mobile agents in farms and warehouses.