Tech | Visa | Scholarship/School | Info Place

AI trains robot dog to walk on yoga ball

You’ve seen cowboys riding bulls at rodeos, and now here’s a robot dog that can balance and walk continuously on a yoga ball!

Watching the four-legged devices move and adapt is fun, but it reinforces how powerful the technology is and how artificial intelligence like GPT-4 can train robots to perform difficult tasks more efficiently than humans.

Dr. Eureka is in charge. It is an open source software package for training robots to perform real-world instructions using LLM (e.g. ChatGPT 4).

Simulated physics are used in a virtual environment, a “simulated reality” system, before tasks are performed in real life.

Jim Fan, one of the developers, said: “The yoga ball task is particularly difficult because the surface of a bouncy ball cannot be accurately simulated. DrEureka, however, has no problem searching vast spaces of simulated to real-life configurations and enables the dog to maneuver a ball over a variety of terrains and even walk sideways! “

How does LLM train robots?

The “Dr” in DrEureka stands for “Domain Randomization”. It represents the randomization of variable factors such as humidity, friction, mass, center of gravity in a simulated situation.

By providing some instructions in the LLM, the AI ​​can process the information to write code and determine a reward/punishment system to teach the robot, in effect, where 0 equals failure and any value above zero is success – but the higher the score, the more Big reward. Framing can be done by setting minimum and maximum breakpoints for variables such as ball bounce, drive, and body movement.

With the help of powerful LLM, it is possible to effortlessly create a large number of parameters for the training system to execute commands simultaneously – let the robot dog do its thing on the yoga ball.

Image source: Ideographies


#trains #robot #dog #walk #yoga #ball

Leave a Reply

Your email address will not be published. Required fields are marked *