This isn’t just a pre-production prototype. It uses mass production automotive-grade quality control and precision robotics manufacturing, with automation exceeding 80%. In addition, IRON’s supply chain is mature, with over 85% shared with XPENG’s EV business.
IRON also supports autonomous charging. It will plug and unplug connectors, locate charging locations, and maintain stable operation, even if a minor equipment fault happens. This ensures safe, reliable charging can happen independently.
A couple of weeks ago, XPENG announced that its robotics business raised over $900 million USD, at a post-money valuation of over $6.3 billion, which is the largest single-round private financing ever recorded in China’s embodied AI industry. The purpose was to ramp up production for commercialization. In an industry where promises are often broken, some questioned the valuation.
I mean, that entire article discussing the robot - and not one mention of what it can do, other than walk itself off the assembly line. Which is certainly an impressive technological feat, since bipedal motion and balance isn’t easy. But it’s also not anything particularly useful. And until these things can do useful jobs, we’re not going to see them move past limited penetration.
It can learn. IRON is equipped with XPENG’s multi-modal LLM model.
*IRON uses AI-driven intelligence to meet high safety standards. * While we have seen many robot demonstrations before, including recently at IFA Berlin, they have tended to rely on a level of remote control or pre-programed motion. From the moment it walks off the production line, IRON operates autonomously.
IRON is humanoid shaped and should be able to do many of the tasks that humans perform. elder care-teacher-babysit-factory work-building sercurity-ultimately policing
And production has started. XPENG is delivering on schedule and ahead of what many expected. Now that the production line is live
Meanwhile Tesla’s Optimus production has just been started. But there are production bugs to be worked out.
Why “should” it be able to do that? It almost certainly doesn’t know how to do any of that now. And there’s no way for it to meaningfully learn.
It would require a massive amount of data to teach IRON to be able to manipulate objects in the real world, and that data doesn’t exist. It probably will take many, many years of data collection before it (or any other humanoid robot) can do anything useful in a factory or school or taking care of the elderly. How on earth is it going to learn how to do that? The data for the forces involved just don’t exist in the volume that we have for text and images. We don’t have a training data set, and it will be too expensive to create one.
Or rather, they’re wagering billions to build robots in the anticipation that the AI will catch up, and they might be wrong. Analogous to Musk’s approach to FSD - he also wagered billions expecting that the “brains” necessary for actual autonomy would exist sometime before 2020, and he was simpy incorrect.
Right now, Tesla’s got workers that are spending full working days just repetitively performing household cleaning tasks like wiping a counter. Other firms are teleoperating robots doing housework and cleaning tasks. Those are all incredibly expensive ways to artisanally handcraft a bespoke set of training data for tasks like “moving a washrag across a counter” or “put away dishes” - specifically because those types of datasets don’t exist for free the way we put virtually the entirety of English language text and a trillion labeled pictures on the internet.
So I would say that Xpeng and Tesla are simply way too early here. If we were to develop the brains that could make these robots useful any time in the near future, they’d be way ahead of competitors. I can see why they might want to take that chance. But right now, the current state of play suggests that there’s not much chance of that actually happening.
Humans have to learn one by one. Not so robots. AI neural networks do the AI learning which is distributed to all robots via over-the-air downloads. Of course that learning is not complete without feedback from robots but that does not mean each and every robot. A handful or a few hundred should be enough. Let’s see what AI has to say for itself.
GoogleAI:
Fleet learning lets AI train an entire group of robots using data from only a small fraction of the machines.
You described fleet learning (also called cloud robotics or swarm intelligence). Your core idea is correct: robots do not need to learn everything on their own.
How Fleet Learning Works
Shared Experience: When a few robots face a new problem in the real world, they record the data.
Cloud Processing: They send this data back to a central cloud computer.
Model Updates: The AI neural network trains on this new information to find a better solution.
Instant Delivery: The system sends the new skill to every other robot through an over-the-air (OTA) download.
Why a Few Hundred Robots Are Enough
Pattern Recognition: Neural networks are good at generalization. They spot patterns from a limited set of examples.
Data Diversity: If a few hundred robots test the code in different places, the AI gathers enough varied data to understand the core problem.
Efficiency: Updating millions of machines at once saves time, energy, and money compared to training each unit separately.
The Limits of This Approach
Rare Situations: A “long-tail” problem (a very rare event) might not happen to those first few hundred robots. If the AI never sees it, it cannot learn it.
Safety Rules: Physical robots can cause damage if an update fails. Engineers must test new software carefully before sending it out to the whole fleet.
Doesn’t address the problem. The AI neural network needs to learn how to do stuff in the real world for the robots to be useful. And there doesn’t exist a dataset sufficiently large and robust to let an AI neural network learn to do that, at least not with the methods we currently use. Which, again, is why companies are paying people to do routine household tasks in a lab environment, either wearing mocap suits or being taped the whole time. Or running these home cleaning services at a massive loss, so they can have their teleoperators running robots and harvesting data in people’s homes.
There is no IRL equivalent to the Common Crawl. No dataset for a neural net to learn how much force gets applied to a paper towel, or the coefficient of friction of a countertop, or a million other pieces of non-visual data that they need to have in order to be able to learn how to physically manipulate objects in the real world the way they have learned to manipulate words or images. These companies have to handcraft that data. Imagine how many decades it would have taken to train a modern LLM to do what it does if it couldn’t rely on previously digitized content, but instead the companies had to hire writers and artists to generate the language and image datasets from scratch!
Should be, but color me skeptical for now. I’ve never seen a video of a humanoid robot doing something I would hire the robot to do. I don’t need a robot that runs fast, for example. Just now I searched for the latest Optimus videos and couldn’t find anything recent. I’m sure if Optimus could do something impressive we’d be seeing it.
In the meantime, I think Amazon might be the actual robotics leader. They have AI robots working in the FC’s right now. They started off using specialized robots which are becoming more general. Tesla is trying to start off with general.
