DeepMind’s new self-improving robot is quick to adapt and learn new skills
They are still fragile, even though they have moved out of the laboratory. The AI can be easily thrown off by slight changes in lighting or the environment. These models must also be trained extensively on specific hardware configurations to carry out useful tasks.
This is in stark contrast with the latest LLMs which have shown a remarkable ability to generalize their skills in a wide range of contexts. This has led to a growing interest in seeing if the underlying technology, an architecture known as a Transformer, could lead to breakthroughs for robotics.
Researchers at DeepMind have shown in new results that an AI based on a transformer called RoboCat is able to not only learn many skills but also switch easily between different robot bodies and acquire new skills faster than usual. It’s also able to speed up its learning process by creating its own data.
Source:
DeepMind’s New Self-Improving Robot Is Quick to Adapt and Learn Fresh Skills