DeepMind researchers have combined AI with a sophisticated robot that can learn new tasks from as few as 100 demonstrations.
RoboCat, a groundbreaking robotic AI agent, learns new tasks from as few as 100 demonstrations, improving its skills through self-generated data.
Despite its name, RoboCat is a robotic arm designed to perform complex tasks like stacking different colored blocks in a particular order. View DeepMind’s demonstrations below.
The robot’s innovative self-improving training cycle represents a significant breakthrough in the field of robotics.
RoboCat utilizes DeepMind’s multimodal model Gato, which can process language, images, and actions across simulated and physical environments.
For RoboCat’s training, researchers curated a massive dataset of image sequences and action sets from various robotic arms performing hundreds of tasks. Following the initial training, RoboCat enters a “self-improvement” cycle, tackling new tasks, leading to further refinement.
The cycle consists of the following steps:
- Collecting 100 to 1000 demonstrations of a new task demonstrated with a robotic arm operated by a human.
- Fine-tuning RoboCat on the new task to create a specialized agent.
- The specialized agent then practices the new task or arm around 10,000 times, which results in the generation of more training data.
- Both the demonstration and self-generated data are then incorporated into RoboCat’s existing dataset.
- Finally, an updated version of RoboCat is trained using the augmented dataset.
This process of continuous training and self-improvement means RoboCat’s dataset is exceptionally diverse.
RoboCat adapts and learns from tasks
Notably, RoboCat has proven to be adaptable, quickly learning to operate new robotic arms, some with different configurations than it was initially trained on.
For example, although RoboCat’s training initially involved arms with two-pronged grippers, it successfully adapted to a more complex arm with a three-fingered gripper.
In one experiment, after observing 1000 human-controlled demonstrations, RoboCat successfully maneuvered a new arm to pick up small gears 86% of the time. It also adapted to solve complex tasks requiring precision and understanding, such as extracting the correct fruit from a bowl and solving a shape-matching puzzle.
RoboCat’s abilities don’t plateau – it grows increasingly capable as it learns.
The initial version of RoboCat succeeded in performing unseen tasks 36% of the time after learning from 500 demonstrations per task, whereas the final version more than doubled its success rate to 74%.
RobotCat moves us one step closer to creating versatile, general-purpose robots. Rapid learning, adaptability, and self-improvement are prerequisites to building intelligent robots that integrate into their environment.
While RobotCat’s Gato model is currently confined to an arm, such AIs will eventually control multiple limbs, sensing and reacting to their environment.