Raphael ThysRobots that learn like humans: they’re here
·
Published in
·
3 min read
·
2 days ago
- -
IMAGE: OpenAI’s DALL·E, via ChatGPT
One of the topics that has captured my imagination in recent months is the way that generative algorithms will soon be widely used in the physical world through the development of robots that, like ourselves, learn by imitating.
The fruits of what has become known as “the OpenAI Mafia”, former employees of the company who have left to set up their own startups, have not been long in coming: Covariant, which emerged from the dismantling of the company’s robotics team due to the lack of data needed to train its robots, is building industrial robots capable of training themselves by imitating videos, which is basically equivalent to a ChatGPT for robots. The idea is to provide robotic mechanisms with ways to learn tasks that work in the same way as large language models (LLMs), and that are powered by observation, allowing them to carry out all kinds of movements or tasks that are physically compatible with their characteristics.
The idea a robot can perform tasks without having to go through a costly and rigid programming process, as was traditionally the case with the eye-catching videos of robots dancing or doing parkour, and instead simply copy any task, opens the door to the idea of replacing many blue-collar tasks with machines capable of working round the clock with unbeatable performance and precision.
Until now, the need to program robots using fixed routines that left no degree of freedom other than grading their movements by continuously reading sensors restricted them to repetitive or routine tasks. Incorporating generative algorithms into the process opens the way for highly flexible robotics that continuously learn or adapt to changing environments and situations that reflect our own skills acquisition.
Adapting a robot to carry out sophisticated manual tasks that we now ask of a human worker — not simply tightening a screw or moving something from one place to another, but assignments of much greater complexity — is not going to happen overnight, but the indications are that the main limitation will be the availability of data for training. In the same way that the evolution of generative algorithms took off when some companies began to use data from the internet to train them, we could soon see companies routinely recording their manual workers in order to…