Hardware really grows up

This already was happening last year, but we can pretty confidently say that in 2026, hardware availability is no longer an issue for the field, no matter where you’re located. As a teaser for our upcoming Standards for Embodiments, we have enough solid vendors in the mobile bi-manipulator space to form a stable supply chain:

• AgileX with cheap QDD arms that are easy to ship worldwide

• I2RT with a complete mobile solution and arms that ship from a US warehouse

• Discover Robotics with complete Aloha-style imitation learning rigs

• Unitree, Booster, and AGIBOT with a mix of bipeds for folks who want legged robots

The most interesting launch, however, has to be the SeeedStudio ReArm. It’s been known for a while that you can purchase a QDD arm for $1499 (HexFellow will sell you one for that price if you’re located in China), but the ReArm is the first to assert this price publicly on a storefront. This pricing shifts the cost landscape a lot: suddenly $2000 UMI grippers are no longer competitive and even the $2700 OpenArm kits are looking expensive.

We have not tested the ReArm yet (as of this writing, it is not yet shipping), but the ability to build a complete data collection cell for under $4000 with worldwide availability should make things scale a lot faster.

From video to teleop…and back again?

We were huge on video-only training pipelines last year and were one of the first teams to assert this possibility publicly. Like good businesspeople, we then immediately started scaling teleoperated data, because that was what the customers were interested in at the time and telling the customers they are wrong rarely results in a good ending.

Somehow, the rest of the industry discovered video-only training throughout the year and subsequently, a lot of people became convinced this was the next big thing. Meanwhile, we’re still pretty bullish on teleoperation (but we’re looking at other modalities as well, stay tuned…), so what’s better?

• Teleoperation still provides the best action quality bar none. Because your training dataset is a one-time investment that goes into every model you ever train, putting down capital to get the best possible quality still makes sense.

• However, it will be at least 18 months before we can say robots are “everywhere”; in the meantime, getting good visual diversity from teleoperation remains a challenge. Video, which is collected in the wild, supplements teleoperation with much greater semantic diversity.

There isn’t really a right answer and in any case, the difference between video and teleop is just the embodiment. In the end, the industry will likely build on cross-embodiment pipelines that can leverage both.

Second-generation physical AI companies emerge

We’ve seen a couple of big fundraising announcements which are part of a second wave of robotics companies. The first generation - Physical Intelligence, Generalist, and DeepMind - blazed the frontiers by showing that end-to-end manipulation learning was possible at all and continue to operate as frontier research labs pushing for full generalization. Second-generation companies, in turn, build on this foundation to target narrower verticals. Two prominent ones are:

• Sunday, targeting household robots with a focus on end-user UX.

• 1X, which has been pivoting around for years and finally pivoted into becoming a full-on physical AI company.

These companies no longer push for abstract “generalization,” which is challenging, hard to monetize, and expensive. Instead, their strategy is to build AI models that are general enough to reliably power their use cases. Additionally, they sidestep the model monetization problem by offering complete solutions - hardware, platform, and intelligence - for a subscription model.

Second-generation companies are also generally more interested in “rogue” data modalities than the original wave. Part of this is because of a desire to be different, part of it certainly is investors asking “what is your moat,” and part of it is simply because the rogue modalities are cheaper and they don’t have as much funding as Physical Intelligence/DeepMind.

Data becomes a lively business

Training data is a funny business to get into. On the one hand, it is a profitable business, but on the other, it could be argued that turning the data into models is even more profitable. We’re pretty good at data, but a big reason is that we understand how data imparts knowledge to models. A few of the other big data shops can do this too - they’ve been working for a while with good teams at Meta and DeepMind and have recently disclosed their capabilities to the rest of the world.

Unfortunately, a side effect of robotics being a hot industry is that everyone wants narrative exposure to robotics, and the easiest way to get it is to become a data company. The result is an explosion of what we call “egoslop” companies - small shops located in second-world countries that will claim to sell you data that is good for training your robotics foundation model. In reality, very few of these teams have any background in robotics, so data quality is usually haphazard.

Our current stance on data is this: data companies are just physical AI companies that choose a different path to profitability, by selling data rather than selling model access. You still need all of the moving pieces that go into a “real” AI company - hardware guys, AI scientists, ops guys - just in a different apportionment.

If you want to know where we draw the line between training-grade data and the egoslop, we wrote it down. Our Standards for Data post lays out exactly what we measure: speed, smoothness, diversity, and whether the operator did one thing at a time and recovered cleanly from a mistake. It’s the bar every demonstration on our platform has to clear.

The interesting part is that scoring against that bar is itself work, and it’s work a lot more people can do than you’d think. We’re about to open it up and let our community score real robot data for quality for the first time. More on that in the coming weeks.

Looking forward - what do we want to see next?

The biggest missing piece in the industry is real closed-loop revenue. Right now, most revenue cycles in physical AI are fueled by VC dollars, which is great for bootstrapping the industry but not sustainable in the long term. With a broad view of the space, we clearly see that all of the moving pieces are in place to create real businesses that generate value in the real world in the next 15 months, and we look forward to seeing robots enter the mainstream.