Chatgpt also manages to design a robot

What are the challenges humanity will face in the future? This question was asked by Cosimo Della Santina, professor, and PhD student Francesco Stella from TU Delft, to ChatGPT. Their intention was to make ChatGPT not just a simple robot, but a real helper. Eventually, they decided to take on the food sourcing challenge, and during a conversation with ChatGPT, they came up with the idea of ​​creating a tomato picking robot . The researchers followed all the decisions made by ChatGPT during the design process. According to Stella, ChatGPT’s suggestions were particularly valuable in the conceptual phase. “ ChatGPT has expanded our knowledge in several areas.

For example, it taught us which crops would be more economically viable to automate .” But ChatGPT also made useful suggestions during the implementation: “ We should use a silicone or rubber gripper to avoid crushing the tomatoes ” and “ the best way to drive the robot is to use a Dynamixel motor “. The result of this collaboration between humans and artificial intelligence was a very efficient robotic arm capable of picking tomatoes.

Cooperation between humans and Large Language Models

Researchers have found the collaborative design process extremely positive and rewarding. However, they have noticed that their role as engineers has shifted towards more technical tasks. On Nature Machine Intelligence, they are exploring the various levels of cooperation between humans and Large Language Models (LLMs), such as ChatGPT. In an extreme scenario, the artificial intelligence would provide all the inputs for the robot’s design and the human would follow them uncritically. In this case, the LLM would act as researcher and engineer, while the human being would take on the role of manager , responsible for specifying the objectives of the design.

However, such an extreme scenario is not yet possible with current Large Language Models. And the question arises whether it is desirable. Indeed , the output of the LLMs could be misleading if not verified or validated. AI bots are programmed to generate the ‘most likely’ answer to a question; therefore, there is a risk of misinformation and bias in the field of robotics ,” says Della Santina. Working with LLMs also raises other important issues, such as plagiarism , traceability and intellectual property. Della Santina, Stella, and Hughes will continue to use the tomato-picking robot in their robotics research. Furthermore, they are continuing to study LLMs to design new robots.

In particular, they are examining how artificial intelligences can autonomously design their own bodies. “An open question for the future of our field is how LLMs can assist robot developers without limiting the creativity and innovation needed to meet 21st century challenges,” concludes Stella.