DeepMind created RoboCat, an AI model that can solve and adapt to diverse tasks utilizing various real-world robots. Unlike prior models, RoboCat can perform a variety of jobs across many robotic incarnations. It was trained using photos and action data from simulated and real-world robotic environments. Initially, researchers gathered 100 to 1,000 demonstrations of human-controlled robotic arm operations. These demos were utilized to fine-tune RoboCat, resulting in a specialized model that performed the task 10,000 times. Researchers increased RoboCat’s training dataset and trained successive versions of the algorithm using data from spin-off models and demonstrations.
RoboCat’s final version was trained on 253 tasks and tested on 141 problem variations in both simulation and real-world scenarios. DeepMind reports task-specific success rates ranging from 13% to 99%, with fewer demos resulting in lower success rates. In several cases, RoboCat displayed the ability to learn new activities with as few as 100 demonstrations.
DeepMind considers RoboCat as a tool for lowering the barrier to tackling new robotics challenges. RoboCat may be fine-tuned for new jobs and collect extra data to increase its performance with a limited number of demos. In the future, the study team hopes to minimize the number of demonstrations required to teach RoboCat a new task to less than ten.
