Oct 17, 2019
Automatic Domain Randomization
Using Artificial Intelligence to Solve Variants of the Same Problem
OpenAI is an exciting company. Earlier this year on July the 24th I wrote about their $1 billion investment from Microsoft, and I have additionally been looking into several of their research articles from a social science perspective. However now I have reason to return to their writing due to a new method called Automatic Domain Randomization (ADR). This is a technique which endlessly generates progressively more difficult environments in simulation.
They test their model with perturbations: a deviation of a system, moving object, or process from its regular or normal state or path, caused by an outside influence.
We could call this: disturbing the algorithm or decision-making process. A crowd favourite in OpenAI’s recent project seems to be the ‘Plush giraffe perturbation’ and occurs at 1:13 in the above video. One commenter on Youtube deftly said:
“Giraffe is my favourite test for any technology. Frankly they should benchmark everything by whether it can giraffe.”
OpenAI has just demonstrated that models trained only in simulation can be used for something quite extraordinary. The manipulation problem of as an example solving the rubriks cube is impressive. Not only solving the problem, additionally being robust to tackle the challenges of a changing environment.
“This frees us from having an accurate model of the real world, and enables the transfer of neural networks learned in simulation to be applied to the real world.”
Automatic domain randomization (ADR) automatically generates a distribution over randomized environments of ever-increasing difficulty.
The combination of ADR with OpenAI’s custom robot platform allows them to solve a Rubik's cube with a humanoid robot hand, which involves both control and state estimation problems. See the full video from OpenAI here summarising their project:
This is an exciting development in the field of artificial intelligence and it will be interesting to see where this new ADR method goes next. It is not unlikely that we will see industrial applications following this progress made by OpenAI. Their post ends with this statement:
“We believe that human-level dexterity is on the path towards building general-purpose robots and we are excited to push forward in this direction.”
I hope you enjoyed this short article. This is day 136 of #500daysofAI. I write one new article about or related to artificial intelligence every day for 500 days.
