DeepMind has created a versatile AI model to control various robotic arm models

DeepMind has developed an artificial intelligence model, RoboCat, which can perform a range of tasks on different models of robotic arms.

Here's What We Know

RoboCat has been trained on images and robot action data collected in both simulations and real life. The researchers first collected between 100 and 1,000 demonstrations of the task. They then trained RoboCat on the task, creating a dedicated "spin-off" model that practiced the action an average of 10,000 times.

Using new data and existing demonstrations, the researchers continually increased the dataset and improved the algorithm.

The final version of RoboCat was trained on 253 tasks and tested on 141 variations of these tasks in simulation and the real world. DeepMind claims that RoboCat learned how to control various robotic arms after observing 1,000 human demonstrations over several hours.

During testing, the success rate varied widely, from 13% in difficult conditions to 99% in simple conditions.

In the future, the research team intends to reduce the number of demonstrations to ten.

Source: TechCrunch