DeepMind’s pricey robot simulator is now available for free

Annual licenses used to cost up to $15,000.
Subscribe to Freethink on Substack for free
Get our favorite new stories right to your inbox every week

AI research lab ​​DeepMind, a sister company of Google, has purchased a powerful robot simulator and made its code freely available online — a move that will help engineers train the AI that could control tomorrow’s advanced robots.

A complex world: We use physical contact to interact with the world around us — we grip a glass of water to get a drink, and move from place to place by having our feet make repeated contact with the ground. 

This seems simple, but it’s incredibly complex — we consider a lot of variables every time we use these processes. You might move slower or take smaller steps when the sidewalk in front of you is icy, for example.

A simulator is a risk-free way to train an AI robot.

Robot simulator: To teach an AI robot how to make these decisions, engineers often use simulators — these are virtual worlds that can effectively replicate the physics of the “real world.” 

Roboticists can create a digital robot, place it in the simulator, and let it learn how to navigate the world risk-free — no need to worry about damaging expensive hardware (or buildings, or people) if the AI makes a decision that results in a fall.

This approach to training an AI is also way faster than letting a robot go straight to learning in the real world, because the AI can quickly encounter a wider variety of environments in a robot simulator, and run scenarios thousands of times. 

The challenge: Despite their importance to the development of AI bots, most of the robot simulators available to engineers have significant drawbacks, according to DeepMind.

Commercial software programs are expensive, hard to understand, and often impossible to modify. Open-source simulators, meanwhile, are usually made by students or academics, and when their developers graduate from school or move on to other projects, maintenance and development suffers. 

“We’ve acquired MuJoCo and are making it freely available for everyone, to support research everywhere.”

DeepMind

You’re welcome: In October 2021, DeepMind purchased MuJoCo (Multi-Joint Dynamics with Contact) — a popular commercial simulator that cost roboticists hundreds or even thousands of dollars annually, depending on the license — and promptly announced plans to open-source it.

“As part of DeepMind’s mission of advancing science, we’ve acquired MuJoCo and are making it freely available for everyone, to support research everywhere,” the company wrote.

On May 23, 2022, DeepMind revealed that people could now access its entire source code on GitHub.

“We hope that colleagues across academia and the [open source software] community benefit from this platform and contribute to the codebase, improving research for everyone,” it wrote.

The big picture: Even the most advanced robot simulator can’t perfectly replicate the physics of the real world, and AIs often have trouble making the leap from navigating a virtual world to controlling an actual robot.

But now that MuJoCo is open source, developers have free access to another tool that could help them work out as many kinks in their AI as possible before putting their robot at risk. Developers are also free to build on, modify, and share the code, potentially jump-starting a high-quality software community. 

We’d love to hear from you! If you have a comment about this article or if you have a tip for a future Freethink story, please email us at [email protected].

Subscribe to Freethink on Substack for free
Get our favorite new stories right to your inbox every week
Related
Three founders look to the future at Freethink’s inaugural Great Progression event
The tech community came together for the launch of the Great Progression event series, curated by Peter Leyden and produced by Freethink.
Why AI today is more toddler than Terminator
In “Raising AI,” author De Kai argues that AIs are more like society’s children than machines under our control.
How proof-of-human tech could save the internet
Sam Altman’s World Network uses iris-scanning Orbs to give people a way to prove that they are people — and not AIs — online.
The AI social network war has begun
A secret prototype, a hardware deal with Jony Ive, and millions of AI images suggest OpenAI is making a play for your social media feed.
Are large language models dyslexic?
Despite outperforming humans at many tasks, multimodal LLMs struggle to read time on a clock — just like many people with dyslexia.
Up Next
Subscribe to Freethink for more great stories