“The video does not include any remote operations,” said 1X, a Norwegian humanoid robot manufacturer. “No computer graphics, no cuts, no video speedup, no scripted trajectory playback. All controlled via neural networks, all autonomous, all at 1X speed.”
This is a humanoid maker whose chips OpenAI invested in last year as part of a $25 million Series A funding round. A subsequent $100 million Series B showed how much of an impact OpenAI's attention was worth. It also demonstrated general excitement about universal humanoid robot workers. This concept always seemed distant in the future, but the last two full nuclear fusions have gone away. age.
The humanoids in 1X seem strangely weak compared to what Tesla, Figure, Sanctuary or Agility do. Eve humanoids have no feet or skilled humanoid hands at this point. It rolls on a pair of powered wheels and balances on a third, smaller wheel at the rear. The hands are basic claws. It looks like it's dressed up for a luge venue and has a small, blinking LED smiley face that gives the impression that it's going to ask for food and cuddle, like a Tamagotchi.
A bipedal version called Neo is in the works for 1X, and the hands are well-detailed. But those bits may not have mattered much in the early, pioneering days of general-purpose robots. Most initial use cases go like this: “Pick that up and put it there” – you barely need fingers that can play the piano to do that. And the main places where they will be placed are warehouses and factories with flat concrete floors, so there will be no need to climb stairs or step on anything.
Moreover, many groups have solved bipedalism and beautiful hand hardware. That's not the main obstacle. The biggest hurdle is getting these machines to learn tasks quickly and then execute them autonomously, as Toyota is doing with desk-mounted robotic arms. When Figure 01 “figured out” how to operate a coffee machine on its own, it was a big deal. When Tesla's Optimus folded his shirt in the video, it was much less impressive when it turned out to be controlled by a human remote operator.
In that context, check out this video from 1X.
Any neural network. Everyone is autonomous. All 1X speed | 1X Studio
The above isn't incredibly complicated or sexy. You can't fold your shirt or the coffee machine won't work. But there are fully formed robots whose job it is to pick up and put down items. Hold it at ankle height and waist height. Put it in a box, trash can, or tray. They pick up toys on the floor and throw them away.
They also open the door on their own, walk to the charging station and plug in. Using what appears to be an unnecessarily complicated squatting motion, the plug is placed near the ankle.
Simply put, these jiggers are doing almost exactly what they were supposed to do in early general-purpose humanoid use cases, trained “purely on data,” according to 1X. Basically, the company trained 30 Eve bots, each on a variety of individual tasks, using imitation learning via video and teleoperation. These learned behaviors were then used to train a “base model” that could perform a wide range of behaviors and actions. That base model was then fine-tuned for environment-specific capabilities, such as warehouse tasks or common door manipulations, and finally trained the bot on the specific tasks it needed to perform.
How logistics develops | 1X Android Eve
This final step is probably the one that will occur on-site at the customer location as the bot is given routine tasks. “Data collection and training on desktop GPUs takes just a few minutes,” says 1X. Perhaps in an ideal world, that would mean someone donning a VR helmet and performing a task for a moment, then deep learning software combines that task with the bot's core capabilities to execute it a few thousand times. If you perform simulations that test different random factors and outcomes, your bot will perform well.
“Over the past year, we've been building a data engine to solve general-purpose mobile manipulation tasks completely end-to-end,” Eric Zhang, vice president of AI at 1X, said in a blog post. “We've convinced ourselves that it works, so now we're “We are hiring AI researchers in the SF Bay Area to scale our robots and remote operators by up to 10x.”
These are pretty neat things. I wonder when it will be ready for prime time.
Source: 1X