AI

DeepMind’s RoboCat learns to perform a range of robotics tasks

Comment

Google DeepMind presented onstage
Image Credits: Google

DeepMind says that it has developed an AI model, called RoboCat, that can perform a range of tasks across different models of robotic arms. That alone isn’t especially novel. But DeepMind claims that the model is the first to be able to solve and adapt to multiple tasks and do so using different, real-world robots.

“We demonstrate that a single large model can solve a diverse set of tasks on multiple real robotic embodiments and can quickly adapt to new tasks and embodiments,” Alex Lee, a research scientist at DeepMind and a co-contributor on the team behind RoboCat, told TechCrunch in an email interview.

RoboCat — which was inspired by Gato, a DeepMind AI model that can analyze and act on text, images and events — was trained on images and actions data collected from robotics both in simulation and real life. The data, Lee says, came from a combination of other robot-controlling models inside of virtual environments, humans controlling robots and previous iterations of RoboCat itself.

To train RoboCat, researchers at DeepMind first collected between 100 to 1,000 demonstrations of a task or robot using a robotic arm controlled by a human. (Think having a robot arm pick up gears or stack blocks.) Then, they fine-tuned RoboCat on the task, creating a specialized “spin-off” model that practiced on the task an average of 10,000 times.

Leveraging both the data generated by the spin-off models and the demonstration data, the researchers continuously grew RoboCat’s training dataset — and trained subsequent new versions of RoboCat.

DeepMind RoboCat
Image Credits: DeepMind

The final version of the RoboCat model was trained on a total of 253 tasks and benchmarked on a set of 141 variations of these tasks, both in simulation and in the real world. DeepMind claims that, after observing 1,000 human-controlled demonstrations collected over several hours, RoboCat learned to operate different robotic arms.

While RoboCat had been trained on four kinds of robots with two-pronged arms, the model was able to adapt to a more complex arm with a three-fingered gripper and twice as many controllable inputs.

Lest RoboCat be heralded as the end-all be-all of robot-controlling AI models, its success rate across tasks varied drastically in DeepMind’s testing — from 13% on the low end to 99% on the high end. That’s with 1,000 demonstrations in the training data; the successes were predictably less common with half as many demonstrations.

Still, in some scenarios, DeepMind claims that RoboCat was able to learn new tasks with as few as 100 demonstrations.

Taken further, Lee believes that RoboCat could herald a lowering of the barrier to solve new tasks in robotics.

“Provided with a limited number of demonstrations for a new task, RoboCat can be fine-tuned to the new tasks and in turn self-generate more data to improve even further,” he added.

Going forward, the research team aims to reduce the number of demonstrations needed to teach RoboCat to complete a new task to fewer than 10.

More TechCrunch

Welcome to Week in Review: TechCrunch’s newsletter recapping the week’s biggest news. This week Apple unveiled new iPad models at its Let Loose event, including a new 13-inch display for…

Why Apple’s ‘Crush’ ad is so misguided

The U.K. Safety Institute, the U.K.’s recently established AI safety body, has released a toolset designed to “strengthen AI safety” by making it easier for industry, research organizations and academia…

U.K. agency releases tools to test AI model safety

AI startup Runway’s second annual AI Film Festival showcased movies that incorporated AI tech in some fashion, from backgrounds to animations.

At the AI Film Festival, humanity triumphed over tech

Rachel Coldicutt is the founder of Careful Industries, which researches the social impact technology has on society.

Women in AI: Rachel Coldicutt researches how technology impacts society

SAP Chief Sustainability Officer Sophia Mendelsohn wants to incentivize companies to be green because it’s profitable, not just because it’s right.

SAP’s chief sustainability officer isn’t interested in getting your company to do the right thing

Here’s what one insider said happened in the days leading up to the layoffs.

Tesla’s profitable Supercharger network is in limbo after Musk axed the entire team

StrictlyVC events deliver exclusive insider content from the Silicon Valley & Global VC scene while creating meaningful connections over cocktails and canapés with leading investors, entrepreneurs and executives. And TechCrunch…

Meesho, a leading e-commerce startup in India, has secured $275 million in a new funding round.

Meesho, an Indian social commerce platform with 150M transacting users, raises $275M

Some Indian government websites have allowed scammers to plant advertisements capable of redirecting visitors to online betting platforms. TechCrunch discovered around four dozen “gov.in” website links associated with Indian states,…

Scammers found planting online betting ads on Indian government websites

Around 550 employees across autonomous vehicle company Motional have been laid off, according to information taken from WARN notice filings and sources at the company.  Earlier this week, TechCrunch reported…

Motional cut about 550 employees, around 40%, in recent restructuring, sources say

The deck included some redacted numbers, but there was still enough data to get a good picture.

Pitch Deck Teardown: Cloudsmith’s $15M Series A deck

The company is describing the event as “a chance to demo some ChatGPT and GPT-4 updates.”

OpenAI’s ChatGPT announcement: What we know so far

Unlike ChatGPT, Claude did not become a new App Store hit.

Anthropic’s Claude sees tepid reception on iOS compared with ChatGPT’s debut

Welcome to Startups Weekly — Haje‘s weekly recap of everything you can’t miss from the world of startups. Sign up here to get it in your inbox every Friday. Look,…

Startups Weekly: Trouble in EV land and Peloton is circling the drain

Scarcely five months after its founding, hard tech startup Layup Parts has landed a $9 million round of financing led by Founders Fund to transform composites manufacturing. Lux Capital and Haystack…

Founders Fund leads financing of composites startup Layup Parts

AI startup Anthropic is changing its policies to allow minors to use its generative AI systems — in certain circumstances, at least.  Announced in a post on the company’s official…

Anthropic now lets kids use its AI tech — within limits

Zeekr’s market hype is noteworthy and may indicate that investors see value in the high-quality, low-price offerings of Chinese automakers.

The buzziest EV IPO of the year is a Chinese automaker

Venture capital has been hit hard by souring macroeconomic conditions over the past few years and it’s not yet clear how the market downturn affected VC fund performance. But recent…

VC fund performance is down sharply — but it may have already hit its lowest point

The person who claims to have 49 million Dell customer records told TechCrunch that he brute-forced an online company portal and scraped customer data, including physical addresses, directly from Dell’s…

Threat actor says he scraped 49M Dell customer addresses before the company found out

The social network has announced an updated version of its app that lets you offer feedback about its algorithmic feed so you can better customize it.

Bluesky now lets you personalize main Discover feed using new controls

Microsoft will launch its own mobile game store in July, the company announced at the Bloomberg Technology Summit on Thursday. Xbox president Sarah Bond shared that the company plans to…

Microsoft is launching its mobile game store in July

Smart ring maker Oura is launching two new features focused on heart health, the company announced on Friday. The first claims to help users get an idea of their cardiovascular…

Oura launches two new heart health features

Keeping up with an industry as fast-moving as AI is a tall order. So until an AI can do it for you, here’s a handy roundup of recent stories in the world…

This Week in AI: OpenAI considers allowing AI porn

Garena is quietly developing new India-themed games even though Free Fire, its biggest title, has still not made a comeback to the country.

Garena is quietly making India-themed games even as Free Fire’s relaunch remains doubtful

The U.S.’ NHTSA has opened a fourth investigation into the Fisker Ocean SUV, spurred by multiple claims of “inadvertent Automatic Emergency Braking.”

Fisker Ocean faces fourth federal safety probe

CoreWeave has formally opened an office in London that will serve as its European headquarters and home to two new data centers.

CoreWeave, a $19B AI compute provider, opens European HQ in London with plans for 2 UK data centers

The Series C funding, which brings its total raise to around $95 million, will go toward mass production of the startup’s inaugural products

AI chip startup DEEPX secures $80M Series C at a $529M valuation 

A dust-up between Evolve Bank & Trust, Mercury and Synapse has led TabaPay to abandon its acquisition plans of troubled banking-as-a-service startup Synapse.

Infighting among fintech players has caused TabaPay to ‘pull out’ from buying bankrupt Synapse

The problem is not the media, but the message.

Apple’s ‘Crush’ ad is disgusting

The Twitter for Android client was “a demo app that Google had created and gave to us,” says Particle co-founder and ex-Twitter employee Sara Beykpour.

Google built some of the first social apps for Android, including Twitter and others