Scientists at the University of Delft in the Netherlands and the Ecole Polytechnique Federale de Lausanne (EPFL) in Switzerland have explored different levels of cooperation between humans and large language models (LLMS) to design and develop a tomato harvesting robot using ChatGPT. The paper was published in the journal Nature Machine Intelligence on July 7.
Cosimo Della Santina, assistant professor at the University of Delft and one of the study leaders, said they wanted ChatGPT to design a really useful robot, so they chose the food supply as an opening. While chatting with ChatGPT, they figured out how to build a tomato-picking robot.
Santina says the output provided by ChatGPT is very valuable. During the design phase, ChatGPT informed them which crops would be most economically valuable in terms of automation. In the implementation phase, ChatGPT also made useful suggestions: make clamps out of silicone or rubber to avoid crushing the tomatoes; Dynamixel motors are the best way to drive robots, etc. Humans and artificial intelligence have worked together to create a robotic arm that can pick tomatoes.
The researchers found that AI and humans cooperate positively and efficiently. In the most extreme case, AI provides all the input for robot design, while humans blindly follow it. In this case, ChatGPT acted as the researcher and engineer, and the human acted as the manager, responsible for defining the design goals.
However, the research team also noted that if the output of LLM such as ChatGPT is not verified, it can be misleading. In addition, the human partnership with LLM brings to light a number of other important issues such as plagiarism, traceability, and intellectual property.
Santina and her colleagues plan to use robot tomato-pickers for their research. They will continue to work on LLM to design new robots, with a focus on the autonomy of AI in designing robots.