Google DeepMind’s ‘RoboCat’ agent enables robots to quickly adapt and learn new skills, whether it be solving puzzles or playing video games. It achieves this by leveraging DeepMind’s Gato model, which is capable of processing language, images and actions in simulated as well as physical environments.
What the researchers found was that it was able to acquire new tasks with as few as 100 demonstrations, thanks to its ability to draw from a large and diverse dataset. They cam up with a five step process to help ‘RoboCat’ train robots to learn new tasks, starting by collecting 100-1000 demonstrations of said task. They then fine-tuned it on this new task by creating a specialized spin-off agent and having it practice an average of 10,000 times to generate more training data. This data is then incorporated into its existing training dataset.
- Powerful Productivity: AMD Ryzen 3 3350U delivers desktop-class performance and amazing battery life in a slim notebook. With Precision Boost, get up...
- Maximized Visuals: See even more on the stunning 15.6" Full HD display with 82.58% screen-to-body, 16:9 aspect ratio and narrow bezels
- Backlit Keyboard and Fingerprint Reader: Biometric fingerprint reader and Windows Hello sign-in options help keep your Acer PC secure
We investigate the agent’s capabilities, with large-scale evaluations both in simulation and on three different real robot embodiments. We find that as we grow and diversify its training data, RoboCat not only shows signs of cross-task transfer, but also becomes more efficient at adapting to new tasks,” said Google.
