This Software Lets One Server Run 64 Factory Robots at Once

Here's the problem. Factory robots that pick, sort and assemble need AI that can look at a scene, follow a spoken instruction like 'move the red box left', and physically act. That kind of AI is far too heavy and power-hungry to live inside the robot itself. So the thinking happens on a nearby server, and the robot just sends pictures and waits for the answer. The catch: the answer has to arrive in milliseconds, or the robot could drop something, hit something, or hurt someone. And servers are expensive, so a factory needs one server to handle many robots at once.

Today's software struggles with that. Most systems are built to run one AI model for one job, or they split the work across separate chips and wait in line. Robion takes a different approach. It splits the AI's two jobs — the 'understand what I'm seeing and hearing' part and the 'move the arm' part — so they run side by side on the same chip instead of one after the other. It also stacks several different AI models onto the same chips, and when things get busy, it serves whichever robot is closest to missing its safety deadline first.

The results are striking. Compared with the most widely used multi-stage serving system, Robion handled 6.7 times more robots, and 1.5 times more than running everything as one big pipeline. In a test with eight different AI models on a single 4-chip server, it kept 64 robots running while meeting 98% of its speed deadlines. Think of a restaurant kitchen that goes from cooking one table's order at a time to serving 64 tables without anyone waiting too long.

The catch: this is a research paper, not a product you can buy. The tests ran in controlled conditions, and real factory floors bring network hiccups, dust, dropped connections and years of safety certification before anything ships. Also, one server failing could stall dozens of robots at once — a new kind of risk factories will need to plan around.