State of the machines

Humanoid robotics is no longer a question of whether the machines can move.

They can.

The real questions now are who is putting robots into people’s hands, who is building them at scale, and who is solving the intelligence that makes them useful.

This week’s events touch all three.

Humanoids in the hands of builders

Fauna announced the launch of its Sprout model, a genuinely encouraging step for getting more robots into the real world and one that highlights a growing gap in Western markets.

Today, the humanoids you most see actively deployed in labs, universities, startups, and online demonstrations are often coming from Chinese companies. Firms like Unitree have made it a priority to get large numbers of units into the hands of researchers and small teams around the world.

Western humanoid makers, by contrast, have largely focused on industrial pilots and enterprise partnerships rather than broad, small-batch distribution.

A clear example of this is how widely the Unitree G1 has spread across research and developer communities globally. That kind of grassroots accessibility simply hasn’t existed from Western humanoid manufacturers.

Sprout could be in a great position to fill that gap in the market.

It is a smaller, lightweight humanoid designed to be safe, approachable, and practical for shared indoor environments. Rather than being built for factory pilots or heavy industrial tasks, Sprout is designed as a developer and research platform that allows teams to experiment with embodied AI, applications, and workflows without needing to build hardware from scratch.

Although the early research edition starts at a relatively high price point, Fauna’s intent is clearly to bring that down with scale:

“The early research edition costs around $50,000, but our goal is to meaningfully bring down the price as we scale.”

Big raises, bigger intent

One of the largest funding events in the Chinese humanoid market just took place.

LimX Dynamics closed a $200 million Series B with participation from JD, NIO Capital, SAIC-backed Shangqi Capital, and Dubai-based Stone Venture.

Alongside the funding, LimX recently unveiled TRON 2, a modular platform that can reconfigure between different form factors, and LimX COSA, its operating system designed for physical-world autonomy.

Here’s a look at LimX’s previous funding history:

View from the Korthos evidence interface. This will be available publicly for tracking events and connections.

In the same period, there were more large raises among Chinese companies.

The Beijing Innovation Center of Humanoid Robotics raised $100 million in its first funding round

X Square Robot raised $140 million in a Series A++ backed by ByteDance and HongShan

This is where the volume is

By the start of 2026, a clear pattern had already emerged in humanoid production.

The majority of companies visibly manufacturing and distributing humanoid platforms at scale are based in China. The top producers by unit output are all Chinese firms, and the number of companies operating in this space is heavily concentrated there.

Another example emerged here, at the end of January, the Beijing Innovation Center of Humanoid Robotics officially opened its pilot manufacturing and validation platform and rolled out its 1,000th customized robot on site.

I explored these production and deployment patterns in more detail here →

Every body needs a brain

Figure AI recently unveiled Helix 02, a unified neural control system that showcased Figure 03 operating across a full kitchen task without stopping, resetting, or switching modes.

If you have been tracking demos, a clear pattern appears. Humanoids look far better than they used to. They run, balance, and move with fluid motion. Sometimes they even appear superhuman. But if you pay closer attention, the same gap shows up every time. Usefulness comes down to intelligence and data.

There are a few very different approaches emerging.

A different way to think about robot intelligence

This week, NVIDIA Research released DreamZero, showing a different way to approach robot intelligence.

Most current systems often called Vision Language Action models train robots by showing them an image, giving them a command, and teaching them to predict the next action. These approaches can work well in familiar environments but tend to struggle when the surroundings change or when tasks require movements the system has not seen before.

DreamZero takes another path. Instead of learning to predict actions, it learns to predict how the world will look after actions are taken. By training a model to imagine future frames of video, it learns how objects move, interact, and respond to forces. It learns physical cause and effect rather than simply learning which action follows which input.

In their evaluation tasks, DreamZero generalized to new tasks significantly better than the leading Vision Language Action baselines tested. It was also able to adapt to a different robot embodiment with only around thirty minutes of demonstration data, where conventional approaches would normally require far more training.

In one demonstration, DreamZero successfully operated tasks like pressing elevator buttons, playing pool, and even fanning burgers - tasks it had never been trained on.

These results come from controlled experimental setups rather than unconstrained real world deployments, but they suggest something important. Learning the physical dynamics of the world through prediction may be a powerful way to build robot intelligence that transfers across environments and across different robot bodies.

The money follows the problem

Full stack builders like Figure AI and LimX Dynamics are building the whole system from hardware through to control models and embodied operating systems.

Others are focused purely on robot intelligence and are seeing enormous capital flow.

Skild AI

2023 seed round about $14.5M

July 2024 Series A $300M at $1.5B valuation

2025 reported mid stage round roughly $500M at around a $4.5B valuation

Early 2026 Series C $1.4B at over $14B valuation

In roughly two years, Skild went from seed stage to one of the highest valued private robotics companies.

Physical Intelligence

March 2024 seed $70M

November 2024 Series A $400M at roughly $2B valuation

November 2025 Series B $600M at roughly $5.6B valuation

Between these, dozens of new companies are forming around data collection, simulation, and model training.

That is where a lot of the effort is now going.