Luca, you recently presented how a robot head can be designed with the help of neural networks as part of the lighthouse project KI.Fabrik of the Munich Institute of Robotics and Machine Intelligence (MIRMI). What is the idea behind it?
I try to use the experience and knowledge gained from working with text generators like ChatGPT to make designing products faster and easier. For this, I rely on a comparable open-source technology. So I don't use Dalle-E and ChatGPT from Open AI, but its free open-source counterpart Dalle-mini/Craiyon. The technology behind it is very similar, but uses a different transformer - "BART" instead of "GPT". About the principle: I use the language understanding of the transformer, so I describe which picture I would like to have. The software reassembles images, i.e. generates them independently. For example, I write: "Show me an astronaut on a horse on the beach!" and the programme makes suggestions. Ultimately, I can also use it to show me chairs, lamps or robot heads, whatever I want.
So you let the software make design suggestions for you ...
Yes, the software makes suggestions and after a few iterations I decide on a version of a chair, for example. For the sake of simplicity, I chose robot heads at the inauguration of the new robotics hall as part of the KI.Fabrik at the Deutsches Museum. The idea is then, and this is where the factory of the future comes in, to have this product produced directly on site. With a robot head, which can be made of plastic, for example, I can pass the design on to a 3-D printer and have it produced. A unique piece, and designed by artificial intelligence.
What is currently still missing in the development of batch size 1 production? It is common for factories to produce very high quantities so that their production "pays off".
I am here at the very beginning of my research, in the third year of my doctorate and working in a field that is very dynamic. One of the challenges of producing a product in batch size 1 is that the design must not become too expensive. So if artificial intelligence develops my product for me in a few minutes, while a designer needs a month to do it, it's not only faster, but also more cost-effective. My goal will be to develop a 3D model from the image that is prepared in such a way that I can put it into 3D printing. Then I would have mapped the entire production chain "end to end" from the idea to the haptics. So the wish would be that I enter what I have in mind online in a sentence like ChatGPT, the AI generates a fairly suitable product for me, which I then adjust individually and then hand over to production. However, I'm not sure yet whether it can be realised that way. I won't know until my doctoral thesis is finished.
Publications by Luca Sacchetto, Chair of Data Processing, TUM: