RoboCat is a new Google DeepMind AI model that can control several robotic arms and perform various tasks. This cutting-edge AI model has an extraordinary ability to perform a wide range of jobs utilising multiple types of robotic arms, establishing a new standard in the field of robotics.

Robots are becoming more common daily as they are increasingly programmed to excel at specific tasks. What if machines are capable of learning new skills and maturing over time? Google DeepMind, the lab behind revolutionary AI software like AlphaGo and AlphaFold, aspires to achieve this. Their latest creation, the RoboCat, is an artificial intelligence agent for robotics that learns from its mistakes and can control several different types of robotic arms.

RoboCat

RoboCat is the first agent that works with many real-world robots and can learn to do various tasks. It can learn new tasks with as few as 100 demonstrations and does so far faster than state-of-the-art alternatives. RoboCat was trained using a massive and diverse dataset of images and actions collected from virtual and actual robots. The information provided examples of various activities, including selecting and rearranging things, building structures, and navigating obstacles. RoboCat might use this information to learn independently of the previously assigned jobs.

Multimodal model

To process language, visuals, and actions in both virtual and real-world settings, we developed the multimodal model Gato (Spanish for "cat"). This model serves as the foundation for RoboCat. To train Gato, the team used a massive dataset featuring images and movements of different robot arms completing hundreds of other jobs.

Researchers then put RoboCat through a "self-improvement" training cycle with a new set of tasks after completing the initial training. Each new activity was learned in five stages: 

• Gather one hundred to one thousand examples of a new task or robot being shown with a human-controlled robotic arm.
• Develop a spin-off agent that is a finely tuned version of RoboCat for this new task/arm.
• On average, the repurposed agent will perform 10,000 repetitions of training on this additional task/arm.
• Add the sample and new data you generate to RoboCat's existing dataset.
• RoboCat should be retrained using the updated dataset.

Robotic arm

Due to its extensive training, RoboCat learned to operate several robotic arms in hours. It was trained on two-pronged grippers but could adapt to a more complicated arm with a three-fingered gripper and twice as many controlled inputs.

Source: Deepmind

RoboCat could direct this new arm dexterously enough to pick up gears successfully 86% of the time after seeing 1000 human-controlled demos in hours. It could adjust to accomplish activities that required accuracy and knowledge, such as removing the correct fruit from a dish and solving a shape-matching problem, with the same level of training. 

Abilities

RoboCat's ability to learn and adapt swiftly could spark a revolution in robotics. Conventional robots are built to perform a specific function, making it challenging to acquire new skills. RoboCat's ability to pick up new jobs with just a few demonstrations gives hope for the future of flexible and adaptable robots. 

Source: Deepmind

It can be put to various future uses, significantly impacting our relationships with robots. RoboCat has the potential to usher in a new era of robotics by making them more user-friendly, efficient, and useful.

Conclusion

In addition to its practical applications, the RoboCat is a remarkable example of the intersection of biology and technology. The robot's inventors analysed the anatomy and behaviour of real cats to make it as realistic as possible, and the robot's movements were based on those of real cats.

However, as with any innovative technology, the RoboCat has yet to be met without apprehension. Some worry that as robot technology improves, they could eventually replace humans in the workforce, leading to a loss of jobs and other economic problems. Furthermore, some warn about dangerous people using robots like the RoboCat for spying or fighting.

Image source: Unsplash