Google's DeepMind has developed the RoboCat AI model, which can control a variety of robots to perform a series of tasks

On June 26, Google’s DeepMind said that the company has developed an artificial intelligence model called RoboCat that can control different robot arms to perform a series of tasks. This alone isn't particularly novel, but DeepMind claims that the model is the first to be able to solve and adapt to a variety of tasks, and to do so using different, real-world robots.

RoboCat was inspired by another DeepMind AI model, Gato, which can analyze and process text, images and events. RoboCat's training data includes images and motion data of simulated and real robots, derived from other robot control models in virtual environments, human-controlled robots, and previous versions of RoboCat itself.

Alex Lee, a research scientist at DeepMind and one of the collaborators on the RoboCat team, said in an email interview with TechCrunch: "We showed that a single large model can be used on multiple real-world models. The robot physically solves diverse tasks and can quickly adapt to new tasks and entities."

IT House noted that in order to train RoboCat, DeepMind researchers first used human-controlled robotic arms, Between 100 and 1000 demonstrations of each task or robot were collected in simulated or real environments. For example, let a robotic arm pick up gears or stack building blocks. They then fine-tuned RoboCat, creating a specialized "derived" model on each task and letting it practice an average of 10,000 times. By leveraging data generated by derived models and demonstration data, researchers continue to expand RoboCat's training data set and train new versions of RoboCat.

The final version of RoboCat was trained on a total of 253 tasks and tested on 141 variations of these tasks, both in simulation and in the real world. DeepMind claims that RoboCat learned to operate different types of robotic arms after observing 1,000 human-controlled demonstrations collected over several hours. While RoboCat has been trained on four robots with two-finger arms, the model was able to adapt to a more complex arm with a three-finger gripper and twice as many controllable inputs.

Despite this, RoboCat's success rates on different tasks varied greatly in DeepMind's tests, ranging from a low of 13% to a high of 99%. This is with 1000 demonstrations in the training data; if the number of demonstrations is halved, the success rate will decrease accordingly. In some cases, though, DeepMind claims RoboCat can learn new tasks by observing just 100 demonstrations.

Alex Lee believes RoboCat might make it easier to solve new tasks. “Given a certain number of demonstrations of a new task, RoboCat can fine-tune to new tasks and self-generate more data to improve further,” he added.

Going forward, the research team aims to reduce the number of demonstrations needed to teach RoboCat to complete new tasks to less than 10.

The above is the detailed content of Google's DeepMind has developed the RoboCat AI model, which can control a variety of robots to perform a series of tasks. For more information, please follow other related articles on the PHP Chinese website!