Researchers from Microsoft published their early findings about ChatGPT this week. In this research, researchers intend to check that if ChatGPT can really think beyond texts and the physical world in order to help with robotic tasks. Research was conducted to determine if people could instruct robots using ChatGPT without knowing programming languages.
After providing ChatGPT with the access to object distance information via an application interface, the researchers investigated ChatGPT’s capacity to generate code, primarily in Python, especially for robotics situations including code generation and the zero-shot planning.
To create the Microsoft logo, Microsoft tested ChatGPT with a robotic arm. The researchers also gave ChatGPT the responsibility of developing an algorithm to guide a drone to a predetermined location without causing any damage to the drone or its surroundings. In addition, they examined ChatGPT’s ability to make real-time decisions about a robot’s movement based on sensor data
The researchers however found that, while ChatGPT displayed some rationality, but many of its actions contradict few axioms, even in the face of reasonable bets, decision-making prompts etc.
They also claimed that occasionally ChatGPT might choose the correct action but for wrong reasons. Thus, in their note, Microsoft acknowledges ChatGPT’s limitations and recommends not applying its output to unchecked robots. Despite ChatGPT’s promising outputs, the researchers cautioned against deploying them without thorough analysis.