https://cooltech-z.blogspot.com/


Deep Mind's latest self-improving robot is quickly

Robots are resolutely stupid despite fast breakthroughs in artificial intelligence. However, recent DeepMind research indicates that the same technology used to generate large language models (LLMs) may also be used to develop more adaptive brains for robotic arms.

Autonomous robots are beginning to leave the lab and into the real world, but they are still fragile. The AI that controls them is easily confused by slight changes in the environment or lighting conditions, and these models require considerable training on certain hardware combinations before they can do useful tasks.

Contrast this with the most recent LLMs, which have shown to be adept at generalising their talents to a wide range of activities, frequently in foreign environments. This has sparked a rising curiosity about whether the underlying technology—a transformer-like architecture—could result in advancements in robotics.

In recent findings, scientists at DeepMind demonstrated that a transformer-based AI named RoboCat can not only master a variety of capabilities but can also easily move between several robotic bodies and pick up new abilities far more quickly than is typical. Most importantly, it can quicken its learning by producing its own training data.

The researchers stated in a blog post that "RoboCat's capability to independently learn skills and rapidly self-improve, especially when applied to different robotic devices, will help pave the way towards a new generation of more helpful, general-purpose robotic agents."

The Gato model, which DeepMind researchers released last month, serves as the foundation for the new AI. It can perform a wide range of jobs, including labelling photos, gaming, and even operating robotic arms. This required training on a broad dataset that included text, graphics, and data on robotic control.

But for RoboCat, the group produced a dataset that was expressly targeted towards robotics problems. They produced tens of thousands of examples of four different robotic arms performing hundreds of various tasks, such correctly selecting the right fruit from a basket or building coloured bricks in the proper order.

These demonstrations were carried out by both teleoperated robotic arms operated by humans and virtual robotic arms controlled by task-specific AI. Then, a single, substantial model was trained using this data.

According to the researchers, one of the key benefits of transformer-based architecture is its capacity to consume significantly more data than earlier types of AI. The same way that LLMs have been able to improve their general language skills is by training with a lot of text. According to the researchers, utilising a variety of various hardware combinations, they were able to develop a "generalist" agent that could handle a wide range of robotics jobs.

Additionally, the researchers demonstrated that the model could learn new tasks by honing it using between 100 and 1,000 demonstrations from a robotic arm that was under human control. There are far fewer demonstrations needed to train on a task than would often be the case, indicating that the model is building on top of more fundamental robotic control abilities rather than from scratch.

"This capability will help accelerate robotics research, as it reduces the need for human-supervised training, and is an important step towards creating a general-purpose robot," the researchers noted in their report.

Most intriguingly, the researchers showed RoboCat's capacity for self-improvement. They developed a number of spin-off models that were focused on particular tasks, and they used these models to produce about 10,000 more examples of the task. These were then used to train a new, more effective version of RoboCat, which was subsequently added to the already-existing dataset.

When given 500 examples of a task that had never been done before, the original RoboCat was only able to finish it successfully 36% of the time. However, this number was more than doubled to 74 percent after numerous iterations of self-improvement and training on new tasks.

The model, it must be said, is still not very good at some issues, with success rates below 50% on a number of them and a score of only 13% on one. However, RoboCat's capacity to successfully complete a wide range of tasks and swiftly pick up new ones shows that more versatile robot brains may not be too far away.