Self-Training RoboCat

DeepMind, the renowned artificial intelligence research lab owned by Google, has revealed a fascinating development in robotics. Their team of over 30 researchers has successfully integrated artificial intelligence with a robot named RoboCat, opening doors to a future world of self-training automatons.

In a paper published on the preprint server arXiv, DeepMind explains that RoboCat employs a virtuous cycle of training. Utilising the same technology found in large language models, the robotic cat can rapidly learn new tasks and enhance its performance by constructing its own performance data.

​

Unlike traditional robots programmed for specific tasks, RoboCat's abilities expand as it learns new tasks, thanks to its growing breadth of experience. The robot learns from a dataset of millions of trajectories from both prior tasks and new self-generated data. As its techniques improve, RoboCat transfers its newly acquired skills to other robots, allowing them to build upon those skills.

DeepMind's breakthrough reduces the need for human-supervised training and significantly accelerates robotics research. RoboCat's success rate in tackling previously unlearned tasks increased over time through self-training, with its performance doubling. The robot's ability to learn quickly with just 100 demonstrations is attributed to drawing from a large and diverse dataset.

While this development marks a significant advancement towards creating a general-purpose robot, the question of whether robots will eventually replace human intervention remains open. The 1921 play "R.U.R.: Rossum's Universal Robots" contemplated such possibilities a century ago, envisioning a future where synthetic humanoids replaced human labour.

However, RoboCats present a friendlier outlook, promising advancements in robotics research while emphasising cooperation between humans and machines. As humorist Will Rogers once quipped, "Letting the cat out of the bag is a whole lot easier than putting it back in."

If you want to know more…, refer to the paper titled "RoboCat: A Self-Improving Foundation Agent for Robotic Manipulation"

​