MIT Engineers Teach Robots to Recover from Errors with 'Common Sense'

Overview

• MIT engineers develop a method enabling robots to adjust to disruptions during household tasks, improving their ability to complete tasks without starting over.
• The approach connects robot motion data with the 'common sense knowledge' of large language models (LLMs), allowing robots to self-correct execution errors.
• Researchers demonstrate the method with a robot trained to scoop and pour marbles, showcasing the robot's ability to recover from being nudged off its path.
• The new algorithm learns to identify what subtask a robot is in, facilitating a dialogue between the robot's actions and the LLM's knowledge of subtasks.
• This advancement could significantly reduce the need for human programming or intervention when robots encounter errors during tasks.