Robots Learn New Tricks: How AI is Revolutionizing Robotics

Artificial intelligence technology could be the missing link to allowing roboticists to quickly deploy robots in unfamiliar conditions, according to researchers from the Massachusetts Institute of Technology.

So-called “generative” AI has gained prominence in recent years as a way to quickly synthesize large amounts of information into more manageable formats. The technology has obvious applications for data analysis and content generation, but MIT researchers uncovered another particularly promising angle, as well: teaching robots.

The Challenge of Teaching Robots

Despite the incredible advances in robotics technology in recent years, teaching robots to navigate new, previously unseen circumstances remains a major hurdle. Digital simulations often fail to translate to the real world, and human-gathered data is time-consuming and costly to obtain.

An Innovative Solution: LucidSim

MIT researchers, however, utilized a system known as LucidSim, which turned thousands of prompts into AI-generated descriptions of real-world conditions. Those were then entered into a mapping system to create, in effect, short training videos.

Success in Training

The method successfully taught a four-legged robot to identify objects, climb stairs, and scale a box at a far more successful clip than conventional simulations — all despite never before seeing any data about them.

The Future of Robotics

Researchers said the technology could lead to entirely new tools and models in what is already the “industrial revolution” of robotics — and potentially allow robots to be virtually trained.

The research was presented at the recent Conference on Robot Learning in Germany.