One of the world’s most advanced humanoid robots has been all play and no work. Boston Dynamics’ Atlas is famous for backflips, parkour, and dance mobs. These require extremely impressive robotic control, but they’re also mostly fun research demos.
Now, six months after the legendary robotics lab unveiled an all-new electric Atlas, they’re showing off more of what it can do. A recent video shows Atlas picking auto parts from one set of shelves and moving them over to another, a job currently handled by factory workers.
Apart from being electric, the new Atlas has a unique way of moving. Its head, upper body, pelvis, and legs swivel independently. So, its head might rotate to face the opposite direction of its legs and torso, Exorcist-style, before the rest of its body twists around to catch up.
The new demo highlights another core change for Atlas. Whereas, in the past, Boston Dynamics meticulously programmed the robot’s most impressive maneuvers, the latest video, by contrast, shows a fully autonomous Atlas at work.
“There are no prescribed or teleoperated movements; all motions are generated autonomously online,” according to a description accompanying the video.
Why release this video now? One, humanoid robots are having something of a moment. And two, ditto for artificial intelligence in robotics. Boston Dynamics led the pack for years, but it didn’t rush Atlas into production for commercial use. Neither has it added significant amounts of AI to the equation. Now, it appears to be interested in both.
Last month, the lab, which is owned by Hyundai, announced a partnership with Toyota Research Institute (TRI) to add artificial intelligence, TRI’s specialty, to Atlas. Alongside pure research, the partnership hopes to make Atlas into a general-purpose humanoid.
It’s an intriguing development. In terms of pure robotics, Atlas is world-class. TRI, meanwhile, is working to develop large behavior models, which are like large language models for robotic movement and manipulation. The idea is that with enough real-world data, AI models like this might develop into a kind robotic brain that doesn’t need to be explicitly programmed for every scenario it might encounter.
Google DeepMind has also been pursuing a similar approach with a vision-language-action model called RT-X and united 33 research labs in an effort to assemble a vast new AI training dataset for robotics. And just last week, a TRI-funded MIT project showed off a new transformer algorithm like the one behind ChatGPT, only designed for robotics.
“Our dream is to have a universal robot brain that you could download and use for your robot without any training at all,” CMU associate professor David Held told TechCrunch. “While we are just in the early stages, we are going to keep pushing hard and hope scaling leads to a breakthrough in robotic policies, like it did with large language models.”
Boston Dynamics isn’t alone in its efforts. If anything, it’s late to the party. A host of companies, many born in the last few years, share the goal of general-purpose humanoids. These include Agility Robotics, Tesla, Figure, and 1X, among others.
In an interview with IEEE Spectrum, Boston Dynamics’ Scott Kuindersma said this may be “one of the most exciting points” in the field’s history. At the same time, he acknowledged there’s a lot of hype out there—and a lot of work still to do. Challenges include collecting enough of the right kind of data and dialing in how best to train robotics algorithms.
That doesn’t mean there won’t be more Boston Dynamics videos out soon. “I want people to be excited about watching for our results, and I want people to trust our results when they see them,” TRI’s Russ Tedrake said in the same interview.
AI-Atlas is just getting started.
Image Credit: Boston Dynamics