You’ve seen the cowboy riding a bull at the rodeo, now here’s a robot dog that’s able to constantly balance and walk on a yoga ball!
It’s fun to see a four-legged device move and adapt but it reinforces the serious potential of the technology and what AI like GPT-4 could look like training robot To perform difficult tasks more effectively than humans.
DrEureka is responsible. It is an open-source software package that is used to train robots to carry out real instructions such as LLM. chatgpt 4.
Before performing tasks in real life, simulated physics is used in a virtual setting, a “sim-to-reality” system.
Jim Fan, one of the developers, said, “Yoga ball work is particularly difficult because it is not possible to accurately simulate the bouncy ball surface. Yet DrEureka has no trouble searching the vast space of sim-to-real configurations, and enables the dog to chase the ball across a variety of terrain, even walking sideways!
We trained a robot dog to balance and walk on top of a yoga ball in a complete simulation, and then transferred the zero-shot to the real world. No fine-tuning. Just works.
I’m excited to announce DrEureka, an LLM agent who writes code to train robot skills in simulations, and more… pic.twitter.com/kuG14LmSOh
– Jim Fan (@DrJimFan) 3 May 2024
How does LLM train a robot?
“Domain Randomization” is depicted in Dr.Eureka’s “Dr.” This is to randomize variables such as wettability, friction, mass, center of gravity in a simulated situation.
By providing some instructions in LLM, AI can process the information to write code that determines the reward/punishment system to teach RobotLiterally, where 0 equals failure, and anything above zero is success – but the higher the score, the greater the reward. Framing can be done by determining minimum and maximum breaking points for variables including ball bounce, driving force, and limb speed.
With powerful LLMs, there is no trouble in creating large amounts of parameters for the training system to execute commands simultaneously – to allow the robot dog to do its work on the yoga ball.
Image Credit: Ideogram
ai-trains-robot-dog-to-walk-on-a-yoga-ball
