In the cowboy arena, we have seen brave cowboys riding on the backs of bulls to showcase their skills. But now, in the world of technology, a robotic dog can balance and walk on a yoga ball, which is truly amazing!

Watching this four-legged device move and adapt to its environment is not only interesting but also highlights the seriousness and immense potential of this technology. Especially when we realize that artificial intelligence (AI) like GPT-4 seems to be more efficient than humans in training robots to perform difficult tasks.

Behind all of this is an open-source software package called "DrEureka." It utilizes large language models (LLMs) such as ChatGPT 4 to train robots to execute real-world instructions.

In the "simulation to reality" system, physical phenomena are first simulated in a virtual environment before applying these skills to real-life situations.

Developer Jim Fan said, "The yoga ball task is particularly challenging because simulating the surface of an elastic ball is difficult. However, DrEureka has no problem searching through a large number of simulated-to-real configurations, allowing the robotic dog to manipulate the ball on various terrains and even achieve smooth walking!"

So, how does LLM train robots?

In "DrEureka," "Dr" stands for "domain randomization." This means that in the simulated environment, variables such as humidity, friction, mass, and center of gravity are randomized.

By providing a series of instructions to the LLM, the AI processes this information and generates corresponding code. These codes establish a reward/punishment system to teach the robot in the virtual environment. A score of 0 represents failure, while any score above zero represents success—the higher the score, the greater the reward. Additionally, various variables, including ball bouncing, driving force, and limb movement, can have minimum and maximum breakpoints.

With the powerful support of LLM, it becomes effortless to create a multitude of parameters for the training system to execute commands simultaneously—this is the secret behind the robotic dog's impressive performance on the yoga ball.