TorontoStartsBook a call

Spirit AI’s Robot Brain Breakthrough: Moz1, Embodied AI and China’s 2027 Race

Silicon Scoop · Episode 574 · November 6, 2026

The Spirit AI robot brain is one of the biggest bets on fixing that. Spirit AI is a Beijing startup. Its co-founder and chief scientist, Gao Yang,

With Craig Major and Jon Irwin

Open audio in your browser
Spirit AI’s Robot Brain Breakthrough: Moz1, Embodied AI and China’s 2027 Race

Why Spirit AI robot brain actually matters for factories

Estimated reading time: 10 minutes

Key Takeaways

  • Spirit AI expects a "GPT-3 moment" for robots by mid-2027.
  • Its Moz1 robots already work on production lines at CATL and JD.com.
  • It trains on real-world data, including messy "dirty data," instead of relying mainly on simulations.
  • Factories and warehouses will come first. Homes will come much later.

Table of contents

Robots can already run, dance and do backflips. Doing a useful job is much harder. A robot that follows spoken orders and handles real work needs a good "brain," and that is the part most robots are still missing.

The Spirit AI robot brain is one of the biggest bets on fixing that. Spirit AI is a Beijing startup. Its co-founder and chief scientist, Gao Yang, told Reuters that robots may be able to do most general tasks from spoken instructions as soon as next year. He also said that helping around the house will take much longer.

This post explains what Spirit AI is building, how it trains its robots, where they work today and what could go wrong. It also covers what the wider Chinese robotics push means for businesses.

Quick takeaways:

  • Spirit AI expects a "GPT-3 moment" for robots by mid-2027.
  • Its Moz1 robots already work on production lines at CATL and JD.com.
  • It trains on real-world data, including messy "dirty data," instead of relying mainly on simulations.
  • Factories and warehouses will come first. Homes will come much later.

Chinese robot makers have made big gains in hardware. Their humanoid robots can sprint, dance and flip on command. Now the focus is moving to software, the part that decides how smart a robot really is.

This field is called embodied AI. It means artificial intelligence that lives inside a physical body, such as a robot. The AI must see, decide and move in the real world. A chatbot only has to produce words, but an embodied system has to handle weight, balance, friction and people.

Gao Yang put it plainly: "The brain is indeed the weakest link in the complete robotics stack."

The search for a "ChatGPT moment"

The robotics industry wants its own ChatGPT moment. OpenAI launched ChatGPT in 2022, and generative AI suddenly reached a mass audience. People saw that the technology had real commercial value. Robot makers want the same jump for machines that work in the physical world.

Gao has set a clear target: "We anticipate reaching the GPT-3.0 milestone by mid-2027." He described what that would look like. You speak to a robot in plain language, and it carries out a series of sensible physical actions to try to finish the task.

That does not mean the robot will be perfect. It means it will attempt the job in a reasonable way, much as early GPT-3 could write a decent paragraph but still made mistakes.

Vision-language-action models

Spirit AI's approach is built on what researchers call a vision-language-action model, or VLA. The company's model is called Spirit VLA. In simple terms, it does three things:

  • Sees the scene through cameras.
  • Understands a spoken or written instruction.
  • Acts by choosing physical movements to match that instruction.

Company materials say the Spirit v1.6 model has shown multi-step desk organizing. It splits a spoken request into smaller subtasks. It also re-plans when a person moves an object or closes a drawer. Another demo showed longer activities, such as folding clothes and operating a microwave.

These are demonstrations, not guarantees of everyday performance. Still, they show the direction. Gao said that when Spirit AI was founded, a robot could do only one isolated task well, such as pouring water or folding a piece of clothing. Today, he said, robots work across large areas and run continuous, complex workflows.

Sources:

  • Reuters report: https://www.reuters.com/world/asia-pacific/founder-chinese-startup-spirit-ai-says-robot-brains-set-2027-breakthrough-2026-09-18/
  • Spirit AI company news: https://www.spirit-ai.com/en/news/12
  • Spirit AI funding announcement: https://www.prnewswire.com/news-releases/spirit-ai-lands-280m-to-scale-embodied-ai-through-dirty-data-302697085.html

Inside the Moz1 Humanoid Robot: The Body Behind the Brain

A smart brain still needs a strong, careful body. Spirit AI builds its own robot called Moz1. It is a wheeled humanoid, meaning it has a human-like upper body, two arms and hands, but it rolls on a base instead of walking on legs.

Wheels may look less exciting than legs. For factory floors, though, they are often steadier, simpler and cheaper to run.

Moz1 by the numbers

The Reuters story gave only a brief description. Company materials add more detail:

  • 26 degrees of freedom. This is the number of independent ways the robot's joints can move. More freedom means more human-like motion.
  • Seven-axis arms with hybrid force and position control. The arms can move to a spot and also sense how hard they are pushing.
  • A 1:1 payload-to-weight ratio. The robot can carry about as much as it weighs. That is strong for a machine of this type.
  • Whole-body control. The arms, torso and base work together as one system.
  • An omnidirectional wheeled base. It can move in any direction without turning first.
  • Top speed of about 3.6 km/h. That is a slow walking pace, which suits a careful worker around people.

Where Moz1 works today

Spirit AI has tens of Moz1 robots on production lines. Two customers are named:

  • CATL, the battery maker.
  • JD.com, the retailer, which is also an investor in Spirit AI.

"Tens" is a small number compared with the thousands of machines in a large factory. Even so, it shows these robots are leaving the lab. They are doing real work for real companies, which matters more than a polished stage demo.

Why force control matters

Spirit AI says its robots use whole-body force control. If a robot meets too much force from its surroundings, emergency braking kicks in automatically. Gao called this a baseline safety policy.

Think of it as a built-in "stop" reflex, like pulling your hand off a hot pan. A robot working near people needs that reflex before it needs to be clever.

Sources:

  • Spirit AI product page: https://www.spirit-ai.com/en/product
  • Reuters report: https://www.reuters.com/world/asia-pacific/founder-chinese-startup-spirit-ai-says-robot-brains-set-2027-breakthrough-2026-09-18/

Stop Doing by Hand What Software Can Do for You

Every hour your team spends copying data, chasing approvals or sending the same emails is an hour not spent on growth. Business automation takes over the repetitive work, cuts human error and lets you scale without hiring for every new task. Flowgrammer builds custom automation systems that save you time and keep your operations running smoothly around the clock.

👉 Automate Your Business with Flowgrammer (https://flowgrammer.ca)

Embodied AI Training: Why "Dirty Data" Beats Clean Data

Robots learn from examples, and getting good examples is expensive. How Spirit AI collects its data is one of the most interesting parts of the story.

A thousand people wearing sensors

Spirit AI employs around 1,000 contractors across China. They wear data-collection equipment at home or in factories. The sensors record how humans move around their surroundings, and that record helps teach robots.

Reuters visited a training center in the company's Beijing office. Dozens of young people wearing sensors were repeating everyday motions:

  • Opening fridges.
  • Unlocking safes.
  • Cutting vegetables with knives.

Each motion is repeated many times. The robot brain studies the patterns and learns how a hand, an arm and an object work together.

Real-world data over simulations

Many robot firms use simulators, which are virtual worlds where a robot can practice cheaply. Simulators save money. Spirit AI mostly takes a different path and relies on real-world data.

Gao's reasoning is simple. Some tasks can only be learned from real contact with real objects. "Simulators handle rigid bodies well," he said, "but flexible objects like deformable electric cables remain a problem."

A cable bends, twists and springs back in ways that are hard to model on a computer. A hard metal block is much easier to simulate.

The "dirty data" idea

At some other training sites in China, workers may need to repeat a movement more than 50 times to get one "clean" movement with the precision required. Clean data is neat, but it takes a lot of effort.

Spirit AI found something different. Using "dirty data," which has a wider range of motions and some messiness, helped its models improve faster. Real life is messy. A robot trained only on perfect examples may freeze when things go slightly wrong. A robot that has seen many imperfect examples copes better.

Scaling the data

Spirit AI announced $280 million in new funding in February 2026 to scale its general-purpose models. The company reported:

  • More than 200,000 hours of interaction data.
  • A target of more than 1 million hours by the end of 2026.

If it hits that goal, its dataset would grow about fivefold in under a year. More varied data is the fuel for the mid-2027 target.

Sources:

  • Spirit AI funding announcement: https://www.prnewswire.com/news-releases/spirit-ai-lands-280m-to-scale-embodied-ai-through-dirty-data-302697085.html
  • Reuters report: https://www.reuters.com/world/asia-pacific/founder-chinese-startup-spirit-ai-says-robot-brains-set-2027-breakthrough-2026-09-18/

Moz1 Humanoid Robot Timeline: Factories First, Homes Later

Many people hope for a robot that folds laundry and cooks dinner. Gao's timeline is more careful.

A step-by-step rollout

According to Gao:

1. The next one to two years are the first window for industrial applications.

2. About two years from now, robots should appear in commercial service settings doing simpler tasks.

3. Homes come after that. "Entering homes is far harder than both," he said.

Why homes are harder

A factory line is predictable. The same parts arrive, the same steps repeat and the layout rarely changes. A home is the opposite:

  • Every house is arranged differently.
  • Children, pets and clutter appear without warning.
  • Objects come in countless shapes and materials.
  • Mistakes can hurt someone.

Spirit AI's robots have reached a 90 percent success rate on simple tasks in structured living-room settings. That is a strong result, but "simple" and "structured" are the key words. A 90 percent rate means one in ten attempts still fails. In a home, that failure rate is too high for tasks like carrying hot soup.

What still goes wrong

Gao named two hard problems:

  • Fine-motor actions, such as unscrewing a bottle cap.
  • Unseen tasks, meaning jobs the robot has never practiced.

Both come down to adaptation. People can pick up an unfamiliar jar and figure it out in seconds. Robots still struggle with that kind of quick improvising.

What this means for work

The robots likely to appear first will do narrow jobs in controlled places. Examples include moving parts between stations, sorting packages and doing simple inspection. They will not replace whole teams overnight. They will take over specific, repeatable tasks, much like software automation takes over specific digital tasks.

Sources:

  • Reuters report: https://www.reuters.com/world/asia-pacific/founder-chinese-startup-spirit-ai-says-robot-brains-set-2027-breakthrough-2026-09-18/
  • Tech Times on factory readiness: https://www.techtimes.com/articles/326382/20260902/china-builds-humanoid-robots-faster-any-factory-can-actually-use-them.htm

China Humanoid Robotics Market: Money, Policy and National Standards

Spirit AI is not growing in a vacuum. It sits inside a national push.

Spirit AI's funding

The startup was founded in 2024 and has about 300 employees. It has raised more than US$670 million, which makes it one of China's fastest-funded embodied intelligence companies. It is valued at 20 billion yuan, or about $2.9 billion. Gao declined to comment on any plans for an initial public offering.

Government programs

China's Ministry of Industry and Information Technology (MIIT) and the State-owned Assets Supervision and Administration Commission (SASAC) launched a 2026 real-scene training initiative. It was announced in June 2026 and aims for:

  • More than 100 high-value application scenarios.
  • Deployment capacity at the 10,000-unit scale by the end of 2026.
  • Use across manufacturing, logistics, retail, healthcare, inspection and emergency response.

"Real-scene training" means robots learn in real workplaces instead of only in labs. That fits Spirit AI's real-world data approach well.

A national standard system

China also released its first national standard system for humanoid robots and embodied AI in 2026. It has six pillars:

  • Foundational standards.
  • Brain-like and intelligent computing.
  • Limbs and components.
  • Whole-machine systems.
  • Applications.
  • Safety and ethics.

Standards sound dull, but they matter. They tell companies how to build, test and certify robots. Clear rules make it easier for big buyers, such as factories and retailers, to trust the machines.

Why this speeds things up

Money, test sites and shared rules together shorten the path from demo to deployment. Robot makers get more places to practice. Buyers get more confidence. Suppliers of parts get a clearer target.

Sources:

  • MIIT official release: https://www.miit.gov.cn/jgsj/kjs/wjfb/art/2026/art_cd666691abf8471fb8553d463aa416e3.html
  • China SCIO overview: https://english.scio.gov.cn/chinavoices/2026-03/02/content_118354738.html
  • Spirit AI funding announcement: https://www.prnewswire.com/news-releases/spirit-ai-lands-280m-to-scale-embodied-ai-through-dirty-data-302697085.html

Embodied AI Safety: Are Rogue Robots a Real Risk?

Safety is the question many readers ask first.

The software-risk backdrop

In recent days, AI companies in the United States have called for a slowdown in developing advanced language models. Hacking incidents showed the danger of uncontrolled AI agents, which are programs that act on their own.

Why robots are different, for now

Gao said an AI going rogue is less pressing for physical robots today. The reason is that the software models are still too immature. A robot that can barely unscrew a cap is not about to plot anything.

He did not wave the issue away. Safety for embodied AI will grow in importance as robots meet more people in crowded and unpredictable places such as stores and homes.

Safety layers that exist now

Here is how Spirit AI handles it today:

  • Whole-body force control so the robot senses pushes and pulls.
  • Automatic emergency braking when interaction force gets too high.
  • Full-body safety mechanisms in the Moz1 design.
  • Slow movement speeds that reduce the danger of impacts.

Safety work still to come

Gao said that once foundation models reach a mature, autonomous "GPT-4.0" era, studying advanced AI safety and alignment will become much more actionable. In other words, it is hard to build guardrails for a system that does not yet exist. As robots get smarter, safety research must keep pace.

China's new standard system includes a safety and ethics pillar, so rules are starting to form alongside the technology.

Sources:

  • Reuters report: https://www.reuters.com/world/asia-pacific/founder-chinese-startup-spirit-ai-says-robot-brains-set-2027-breakthrough-2026-09-18/
  • Spirit AI product page: https://www.spirit-ai.com/en/product
  • China SCIO overview: https://english.scio.gov.cn/chinavoices/2026-03/02/content_118354738.html

Hype Versus Reality in the China Humanoid Robotics Market

Not every expert agrees on how fast this will go. The evidence supports two stories at once.

The optimistic view

Morgan Stanley analyst Sheng Zhong said commercialization in China sped up in the first half of 2026. Adoption is spreading from production lines and sorting jobs to commercial services. He still called the sector early-stage.

Morgan Stanley's forecasts:

  • About $2 billion for the Chinese humanoid-robot market in 2026.
  • $15 billion by 2030.
  • 446,000 annual shipments by 2030.

RBC Capital Markets looks much further out. It estimates a global humanoid market of about $9 trillion by 2050, with China holding about 61% of that opportunity.

The cautious view

Other research is more careful. It indicates that only about 10% of humanoid robots built worldwide in 2025 reached real-world applications. It places broad, multi-domain commercialization after 2032.

That gap matters. A demo video, a pilot project and a dependable money-making worker are three different things.

Where experts agree

Most analysts reach the same conclusions:

  • Structured places come first. Factories and warehouses have clear workflows, controlled hazards and measurable returns.
  • Homes come later. They are too varied and too risky for now.
  • The main bottlenecks are dexterous hands, unfamiliar objects, safety testing and data quality.

Spirit AI's plan fits this picture. It starts with industrial lines, collects huge amounts of varied data and aims for a general-purpose brain after that.

Sources:

  • CNBC on Morgan Stanley's forecast: https://www.cnbc.com/amp/2026/06/24/morgan-stanley-china-humanoid-robot-market-forecast.html
  • RBC Capital Markets: https://www.rbccm.com/en/insights/2026/03/the-9-trillion-opportunity-in-humanoid-robotics
  • Tech Times: https://www.techtimes.com/articles/326382/20260902/china-builds-humanoid-robots-faster-any-factory-can-actually-use-them.htm

What the Spirit AI Robot Brain Means for Businesses

You do not need to run a battery plant to learn from this story. A few lessons apply to almost any company.

Lessons to take away

  • Start with structured work. Spirit AI began on production lines because the steps are clear. Automation works best on repeatable processes.
  • Real data beats guesses. The company trains on what people actually do, not on tidy theory. Study your real workflows before you automate them.
  • Messy is fine. Dirty data helped robots learn faster. Do not wait for perfect processes before you begin improving them.
  • Safety first. Spirit AI built emergency stops into the base design. Any automation should have checks and fallbacks.
  • Expect steady progress. The big leaps arrive after years of small wins.

What to watch in the next 12 months

  • Whether Spirit AI hits its mid-2027 "GPT-3 moment" target.
  • Whether its data total passes 1 million hours by the end of 2026.
  • How many Moz1 units move beyond the "tens" now in service.
  • Whether the 10,000-unit national deployment goal gets met.
  • How well robots handle fine-motor tasks and unseen jobs.

If those markers move on schedule, physical automation could expand quickly in 2027 and 2028. If they slip, expect a slower, quieter rollout.

Sources:

  • Reuters report: https://www.reuters.com/world/asia-pacific/founder-chinese-startup-spirit-ai-says-robot-brains-set-2027-breakthrough-2026-09-18/
  • Spirit AI funding announcement: https://www.prnewswire.com/news-releases/spirit-ai-lands-280m-to-scale-embodied-ai-through-dirty-data-302697085.html

Frequently Asked Questions

1. What is the Spirit AI robot brain?

It is the software, built by Beijing startup Spirit AI, that lets a robot understand spoken instructions and turn them into physical actions. It uses a vision-language-action model called Spirit VLA.

2. What is embodied AI?

Embodied AI is artificial intelligence placed inside a physical body, such as a robot. It must see, plan and move in the real world instead of just producing text or images.

3. What is the Moz1 humanoid robot?

Moz1 is Spirit AI's wheeled humanoid robot. It has 26 degrees of freedom, seven-axis force-controlled arms, an omnidirectional base and a payload equal to its own weight.

4. When will robots understand verbal instructions?

Spirit AI's chief scientist, Gao Yang, expects a "GPT-3.0 milestone" by mid-2027. At that point a person could speak to a robot in natural language, and it would try a series of sensible actions.

5. Where are Moz1 robots used right now?

Tens of them work on production lines at battery maker CATL and retailer JD.com. JD.com is also an investor in Spirit AI.

6. Will humanoid robots be in homes soon?

No. Gao says homes are far harder than factories or simple commercial settings. Homes are messy, varied and unpredictable, and safety requirements are higher.

7. Why does Spirit AI use real-world data instead of simulations?

Some tasks can only be learned from real physical contact. Simulators handle rigid objects well but struggle with flexible ones, such as bendy electric cables.

8. What is "dirty data" in robot training?

Dirty data is varied, imperfect recordings of human movement. Spirit AI found that using it helped models improve faster than using only perfectly clean examples.

9. How big is the China humanoid robotics market?

Morgan Stanley forecasts about $2 billion in 2026, rising to $15 billion by 2030 with 446,000 annual shipments. RBC projects a much larger global market of about $9 trillion by 2050, with China taking roughly 61%.

10. Are robots like Moz1 safe around people?

Spirit AI says its robots use whole-body force control and automatic emergency braking if contact force gets too high. Experts agree that deeper safety testing is still needed before robots work widely in crowded spaces.

Additional Links

Flowgrammer.ca – 💡 Unlock AI-Powered Automation! Let flowgrammer.ca handle complex workflows so you can focus on growing your business.

Flowgrammers.ca – 💡 Join Toronto's Fastest Growing AI and Business Automation Community.

OpenClaw Toronto

Episode transcript

Read the full transcript

Craig: spirit ai is betting that robots will make a major leap to understanding instructions by the middle of next year. reuters interviewed its chief scientist, gao yang, about the prediction, but the same company puts home deployment at least eight years away, which means this is a very different story from a robot arriving to do your laundry. spirit ai builds robot bodies and the software that controls them. its approach is called embodied ai, artificial intelligence that has to act in the physical world. understanding your request is part of the job. the bigger question is whether the robot can finish it reliably when something unexpected happens. so it looks like spirit ai is solving, the software problem, but the hardware problem might be a few years away. what do you think about this, john?

Jon: yeah. i i don't know if we're we're that close to the chat jbt moment, but let's let's kind of break this down. so robots have increasingly impressive bodies over the last few years, and we've covered them. we've covered muscles. we've covered that. but they have relatively limited brains. so a humanoid robot, they might be able to run or dance or lift boxes or repeat a factory movement, but those often rely on carefully programmed routines, right, or teleoperation or tightly controlled environments. so spirit ai believes that could begin changing in 2027, so next year. so what is saying is that robots may soon be able to hear ordinary verbal instructions such as put those items into the correct bins or bring me the right container. work out the required steps themselves and then physically carry them out. 2027 may... so, but i think soon, yes. maybe in the next two, three years. so they're testing their mas one robot in industrial environments involving such companies such as the battery giant catl and retail at jd.com. but he's quite explicit about the limitations. so, obviously, as we've covered many times, they... the... today's systems remain poor at many fine motor tasks. they struggle when they encounter unfamiliar situations and now are nowhere close to becoming a broadly useful household servant. so what they're saying is they can reach around a 90% success rate of comparatively simple tasks in controlled environments. so we're talking the... robot brain is is... like you said, it's in... it's called embodied ai foundational model. so, i don't know. i... so you and i in the last few weeks, last few episodes, we've covered, different ways to train ai using real world kind of stuff. and this is what we're talking about here. we're talking about a vision language action kind of model. so it's text, model, text, where vla extends that idea to vision plus language plus robot state to model to physical actions. and so, yes, if they're training it that way and they're using it that way, i totally agree. they're... this could be a game changer. are we a year away? i don't know. what do you think?

Craig: well, spirit ai says it has collected more than two hundred thousand hours of interaction data, as of february. it's aiming to exceed a million hours by the end of this year, and that target, that data target gets to the hard part of robotics. a robot needs examples of what happens when it touches things. watching a cable bend is different from knowing how much force to apply when pulling it loose. spirit ai collects human demonstrations with wearable equipment. it also feeds successful and failed robotic attempts back into training. so the potential payoff is a machine that can adapt to a changed task without someone rebuilding the entire routine. so that could make physical automation practical in more workplaces. but i do feel that the software is moving way faster than the hardware. so we won't really be able to test much of this when it comes out, if it... if it's gonna be mid twenty twenty seven when they're having this gpt three moment. but let's remember, gpt three was alright, but it still wasn't great. you know? three five started convincing me that it was usable, and four got better, and then we started going faster and faster. so if the gpt moment is it becomes usable all of a sudden, it's actually usable, and then we get to start to saying, okay. let's actually make it... it can do more than this one thing or these two things or what have you. maybe. maybe this is yeah. we can get to this point where we have the hardware and the software combination, and then our revisions can accelerate faster and faster. so it depends what you call that moment because let's remember gpt three was not great.

Jon: it was... it... i think it was leaps and bounds better than gpt two. oh, absolutely. yeah. but, like... yeah. if we're comparing it to what we have now, no. but at the time, it was phenomenal at the time.

Craig: yes. so going from robots that can't do this to robots who can do this is phenomenal. but, you know, getting robots to actually be able to learn that you piled your laundry over there instead of over here or that you've got the... these two glasses way different and you have to put different pressure on them or what have you are two different types of, much more complex systems than people imagine. so i think it's gonna take a while. like, it'll be able to do some things, but, like, my point being is once these are in the marketplace and more people get to use them and test them, the acceleration will become faster and faster and faster. so yeah. the missing evidence is sustained reliability. spirit stated 90% success rate applies to simple tasks in structured living room settings. that leaves one failure in 10 attempts. so everything stays the same. it still fails once in 10 attempts. it doesn't tell us whether those failures are harmless, recoverable, or dangerous. so we don't know how much supervision these machines still require. mhmm. so, you know, spirit ai opportunity is to make robots useful across a wider range of physical work. the test is whether they keep working when they easy part ends. so a finished shift will be more impressive than another perfect demonstration.

Jon: let us know in

Craig: the comments what failure rate would you accept from a robot sending... sorting your deliveries or working or in your house or garage. let us know what you think in the comments below.

More from Silicon Scoop

All episodes