Microsoft says it has used the聽natural language AI ChatGPT to control a range of robots with simple text commands. The approach means people with no engineering or coding experience will be able to instruct sophisticated robots to carry out tasks, but an expert warns that it could be risky to ask AI models to control robots on behalf of people.
ChatGPT was created by , an AI company that is itself partly owned by Microsoft. The chatbot was trained on a vast amount of data, including source code. It has previously demonstrated the ability to write software based on text prompts and even to fix errors in existing code. Another Microsoft-owned company, GitHub, has also launched a commercial tool that helps programmers to create software using AI.
Microsoft has now used the AI to take text commands from humans and interpret them into code that can control robots directly. In one test, ChatGPT wrote code to instruct a robotic arm to create a Microsoft logo from wooden blocks. In another, it steered a drone to reach a waypoint while avoiding obstacles.
Advertisement
Microsoft said in a blog post that the research was intended 鈥溾��.
Robots are typically controlled by software that has been written by humans and sets out precise instructions, or else by some form of relatively simple neural network AI that can learn to carry out tasks based on large numbers of examples.
Neither Microsoft nor OpenAI responded to a request for comment, but Microsoft said in its blog post聽that ChatGPT allows a user to give a robot instructions and then either review the code it outputs to check for errors or watch a simulated robot executing them. The user can then give feedback if needed.
ChatGPT isn鈥檛 in direct, real-time control of the robot, but simply creates the code that controls it. Microsoft said in its blog post that ChatGPT output shouldn鈥檛 be deployed directly onto a robot without 鈥渃areful analysis鈥�, but experts warn that it is risky to even begin thinking about placing AI in control of physical machines.
at the University of Vienna, Austria, says that regulations may be needed in order to set out where AI can be used, and who is responsible should things go wrong.
鈥淚t鈥檚 a very dangerous thing to just say 鈥榣et鈥檚 give control of these computers to AI鈥�,鈥� says Coeckelbergh. 鈥淭he problem with contemporary artificial intelligence is that it鈥檚 not transparent to the user. That鈥檚 a huge problem.鈥�
Others were sceptical about the ability of language models to program robots for complex tasks. at De Montfort University Leicester, UK, says that AI language models are convincing mimics, but that their real ability 鈥� including their adaptability to control robots 鈥� is often overblown.
鈥淚 think Microsoft, and most people who design robots, over-inflate what they can and can鈥檛 do,鈥� she says.
