– Home robots have struggled post-Roomba due to pricing, practicality, form factor, and mapping issues
– Big companies have resources to address industrial-level robot problems, but consumers can’t easily program or fix robot mistakes
– MIT research uses LLMs to help robots learn from mistakes and self-correct tasks, improving functionality in unstructured environments like homes.
Home robots have struggled post-Roomba due to issues such as pricing, practicality, form factor, and mapping. Even when these issues are addressed, mistakes can still occur. Unlike industrial robots, consumers cannot be expected to learn to program or hire someone to fix errors that arise. MIT researchers are addressing this issue by using large language models (LLMs) to help robots learn from mistakes.
A study to be presented at the International Conference on Learning Representations (ICLR) focuses on incorporating “common sense” into the process of correcting mistakes made by robots. Robots are good mimics but may struggle to adjust to unexpected situations. Traditionally, robots exhaust pre-programmed options before requiring human intervention, which can be problematic in unstructured environments like homes.
Researchers are breaking demonstrations into smaller subsets, rather than treating them as part of a continuous action, to address the limitations of imitation learning in home robotics. LLMs eliminate the need for programmers to label and assign numerous subactions, enabling robots to automatically know what stage of a task they are in and recover on their own when faced with obstacles.
The study involves training a robot to scoop marbles and pour them into a bowl, a simple task for humans but a combination of small tasks for robots. The system responds to sabotage in the demonstrations by self-correcting small tasks, rather than starting from scratch, eliminating the need for human intervention to fix mistakes. This method aims to help robots learn from errors and improve their performance without human assistance.