Poems, Essays, and Books: ChatGPT Explores Designing a Tomato-Harvesting Robot
In a groundbreaking study published in Nature Machine Intelligence, researchers from TU Delft and EPFL delve into the capabilities of ChatGPT, an innovative open AI platform. The team aimed to determine whether ChatGPT could design a useful robot and the implications of this collaboration.
The Quest for Future Challenges
To kickstart their investigation, Cosimo Della Santina and Francesco Stella from TU Delft, along with Josie Hughes from EPFL, prompted ChatGPT with a crucial question: “What are the greatest future challenges for humanity?” ChatGPT, AI’s virtual assistant, steered the conversation towards designing a robot that addresses a specific problem. The team settled on finding a solution for optimizing food supply and ultimately conceived the idea of a tomato-harvesting robot.
Beneficial Suggestions for Design
Working in tandem with ChatGPT proved immensely valuable for the researchers, especially during the concept development phase. Stella emphasized that ChatGPT expanded their knowledge by providing insights from diverse areas of expertise. For instance, the AI assistant suggested automating the most economically valuable crop. Furthermore, ChatGPT offered practical recommendations during the implementation stage, such as using a silicone or rubber gripper to prevent tomato damage and employing a Dynamixel motor for optimal performance. This collaborative effort culminated in the creation of a successful robotic arm designed specifically for tomato harvesting.
ChatGPT as a Researcher
Reflecting on their experience, the researchers noted that the collaborative process with ChatGPT proved to be positive and enriching. Stella acknowledged that their roles transitioned to more technical tasks as engineers. In their published paper, the researchers discuss the varying degrees of cooperation between humans and Large Language Models (LLM), with ChatGPT being one of them. In an extreme scenario, AI becomes the sole provider of input for robot design, with humans blindly following instructions. Here, the LLM assumes the role of a researcher and engineer, while the human acts as a manager, responsible for clarifying design objectives.
Risk of Misinformation and Bias
Currently, LLMs like ChatGPT do not possess the capability for such extreme levels of control. It is essential to consider the desirability of this scenario as well. Della Santina warns that unverified or unvalidated output from LLMs can be misleading. AI bots, by nature, generate ‘most probable’ answers, leading to potential misinformation and bias in the field of robotics. Working with LLMs also raises concerns about plagiarism, traceability, and intellectual property.
The Future of Robotics and LLMs
Della Santina, Stella, and Hughes intend to continue their research on robotics using the tomato-harvesting robot as a tool. They also plan to further explore the use of LLMs in designing new robots. Specifically, they aim to investigate the autonomy of AI in creating their physical forms. Stella concludes by emphasizing the need to strike a balance between utilizing LLMs to assist robot developers while fostering the necessary creativity and innovation demanded by the robotics industry in the 21st century.