DeepMind Unveils Self-Training RoboCat
Someone who remains unidentified and holds an appreciation for cats once expressed, “Cats and computers share a common trait—they both exert dominance over the Internet.“
Researchers at Google’s DeepMind recently combined artificial intelligence and a robot named RoboCat. Although it doesn’t currently dominate the Internet, it is anticipated to take a significant leap into a future where self-training automatons prevail.
DeepMind’s team, consisting of over 30 researchers, employed the same technology used in large language models. They announced a major breakthrough with RoboCat, a robotic cat that not only quickly learns new tasks but also enhances its performance by generating its own performance data.
In a paper posted on the preprint server arXiv, DeepMind stated, “RoboCat follows a virtuous cycle of training. The more new tasks it acquires, the more adept it becomes at learning additional new tasks.“
Until now, robots were typically programmed to perform specific predefined tasks. However, with the advent of large language models, the skill sets of robots began to expand. Nevertheless, training them on the vast amounts of data required immense time investment.
DeepMind claimed that RoboCat exhibits a remarkable ability to swiftly acquire new tasks, such as correctly fitting differently shaped puzzle pieces into corresponding slots or arranging fruit in a bowl. By leveraging a vast dataset comprising millions of previous task trajectories and self-generated data, the robot was able to advance and tackle more complex tasks.
The researchers stated that RoboCat’s progress and enhanced capabilities stemmed from its increasing breadth of experience, analogous to how individuals acquire a diverse range of skills as they deepen their knowledge in a particular field. As RoboCat refined its techniques, the newly acquired behaviors were transferred to other robots, who further built upon those skills.
Human-controlled robotic arm
To fine-tune its performance, the robot received between 100 and 1,000 demonstrations from a human-controlled robotic arm. Specific task-oriented models were then trained using the obtained data, which was subsequently incorporated into the general instruction pool.
Although RoboCat initially achieved a 36% success rate when tackling unfamiliar tasks, its performance improved significantly over time. Through self-training, its success rate doubled, showcasing the effectiveness of the learning process.
DeepMind researchers expressed that RoboCat learns at a significantly faster pace compared to other cutting-edge models. With access to a large and diverse dataset, it can grasp a new task after as few as 100 demonstrations.
This advancement is considered a significant stride in expediting robotics research by reducing the reliance on human-supervised training. It also represents a crucial step toward developing a versatile and multipurpose robot.
The research paper titled “RoboCat: A Self-Improving Foundation Agent for Robotic Manipulation” was published on June 20.
Does this mean that such robots will eventually eliminate the need for human involvement?
This question was contemplated a century ago in the 1921 play “R.U.R.: Rossum’s Universal Robots” by Czech writer Karel Čapek. The play envisioned a factory that produced synthetic humanoids capable of continuous work, ultimately leading to an 80% reduction in labor costs. It was in this play that the term “robot” was first used, derived from the Czech word “robota,” which denoted “forced labor by serfs.”
Ultimately, the robots revolted and eradicated the human race.
On a more positive note, we can anticipate that RoboCats will exhibit a friendlier disposition.
However, it is important to bear in mind the wise words of humorist Will Rogers, who once remarked, “Revealing a secret or truth is much simpler than attempting to conceal or undo it.”
Read the original article on PHYS.
Read more: Researchers Create a New Robotic Bee with Full Freedom of Movement.
