Shulou(Shulou.com)11/24 Report--

CTOnews.com, June 26 (Xinhua)-- Google's DeepMind said that the company has developed an artificial intelligence model called RoboCat, which can control different robotic arms to perform a series of tasks. This alone is not particularly novel, but DeepMind claims that the model is the first to solve and adapt to multiple tasks, using different, real-world robots.

RoboCat is inspired by Gato, another AI model of DeepMind, which can analyze and process text, images, and events. The training data of RoboCat includes images and motion data of simulated and real robots, which come from other robot control models in the virtual environment, human-controlled robots and previous versions of RoboCat itself.

"We have demonstrated that a single large model can solve a variety of tasks on multiple real robot entities and can quickly adapt to new tasks and entities," Alex Lee, a research scientist at DeepMind and a collaborator on the RoboCat team, said in an email interview with TechCrunch.

CTOnews.com noted that in order to train RoboCat,DeepMind researchers to first use a human-controlled robotic arm, 100 to 1000 demonstrations of each task or robot were collected in a simulated or real environment. For example, let the robotic arm pick up gears or stack building blocks, and so on. They then fine-tuned RoboCat to create a special "derivative" model for each task, allowing it to practice an average of 10000 times. By using the data and demonstration data generated by the derivative model, the researchers continue to expand the training data set of RoboCat and train a new version of RoboCat.

The final version of RoboCat was trained on a total of 253 tasks in both the simulated and real world, and tested on 141 variants of these tasks. DeepMind claims that after observing 1000 demonstrations of human control collected in a few hours, RoboCat learned to operate different types of robotic arms. Although RoboCat has been trained on four robots with two-claw arms, the model can adapt to a more complex manipulator with a three-finger fixture and twice the controllable input.

Nevertheless, the success rate of RoboCat in DeepMind tests varied greatly from the lowest 13% to the highest 99% for different tasks. This is when there are 1000 demonstrations in the training data; if the number of demonstrations is halved, the success rate will be reduced accordingly. In some cases, however, DeepMind claims that RoboCat only needs to observe 100 demos to learn new tasks.

Alex Lee believes that RoboCat may make it easier to solve new tasks. "as long as a certain number of new task demos are given, RoboCat can fine-tune to the new task and can generate more data to further improve it." He added.

In the future, the goal of the research team is to reduce the number of presentations needed to teach RoboCat to complete new tasks to less than 10.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.