Chinese Robotics Firm Says Humanoids Could Follow Verbal Instructions For Most Tasks By 2027,...

Humanoid robots could become capable of carrying out most general-purpose tasks from verbal instructions as early as next year, but putting machines to work reliably inside homes will take significantly longer, according to the co-founder and chief scientist of Chinese robotics company Spirit AI.

The prediction comes as China's humanoid robotics industry shifts its focus from increasingly capable hardware to the software systems that give robots the ability to understand instructions, make decisions and execute sequences of physical actions.

Chinese humanoid robots have recently demonstrated advanced physical abilities, including sprinting, dancing, and performing backflips. The next challenge is turning those demonstrations into machines capable of performing economically useful work across a broad range of environments.

That field, commonly known as "embodied AI," is emerging as one of the most closely watched areas of robotics development.

"The brain is indeed the weakest link in the complete robotics stack," Gao Yang, Spirit AI's co-founder and chief scientist, told Reuters at the company's Beijing offices on Thursday.

For the robotics industry, the objective is to achieve what some executives describe as a "ChatGPT moment," a software breakthrough that makes sophisticated robotic systems useful to a much broader market.

OpenAI's launch of ChatGPT in 2022 demonstrated how quickly an advanced AI technology could move from research laboratories into mainstream consumer and commercial use. Robotics companies are now looking for an equivalent breakthrough that would allow robots to move beyond highly controlled demonstrations and individual industrial tasks.

Spirit AI expects that transition to begin with natural-language interaction.

"We anticipate reaching the GPT-3.0 milestone by mid-2027. You will be able to speak to a robot in natural language, and it will execute a series of reasonable physical actions to attempt the task," Gao said.

The prediction does not mean robots will be capable of reliably performing every household activity by then. Gao expects industrial applications to develop first, followed by simpler commercial services, with domestic environments presenting the greatest challenge.

"The next one to two years mark the initial window for industrial applications. Two years from now, we'll see robots deployed in commercial service settings doing simpler tasks. Entering homes is far harder than both," said Gao, who is also an assistant professor of robotics at Tsinghua University.

The difference is largely about the complexity and unpredictability of physical environments. A factory production line can be structured around a relatively narrow set of tasks, while homes contain an almost unlimited range of objects, layouts and unexpected situations.

Spirit AI currently has tens of its Moz1 wheeled humanoid robots deployed on production lines at battery manufacturer CATL and retailer JD.com, which is also an investor. The 300-person startup has raised more than $670 million since its founding in 2024 and is currently valued at about 20 billion yuan, or $2.9 billion. Gao declined to comment on whether the company plans to pursue an initial public offering.

Building the Robotic "Brain"

Spirit AI is investing heavily in data collection to improve the software controlling its robots. The company employs about 1,000 contractors across China who use data-collection equipment in homes and factories to record how humans interact with physical environments.

At a training center in Spirit AI's Beijing office, Reuters reported dozens of workers equipped with sensors repeatedly performing everyday actions, including opening refrigerators, unlocking safes, and cutting vegetables with knives. The objective is to provide AI systems with examples of how people manipulate objects and move through different environments, creating the training data needed to make robots more adaptable.

Spirit AI said its robots have achieved a 90% success rate on simple tasks in structured living-room environments. The company nevertheless faces substantial difficulties when robots encounter unfamiliar situations or require precise manipulation.

Tasks such as unscrewing a bottle cap can require fine motor control that remains difficult for current systems. Robots also struggle when confronted with objects or tasks that were not represented sufficiently in their training data.

Gao said Spirit AI relies heavily on real-world data rather than virtual simulations. Many robotics companies use simulated environments to generate training data at lower cost, but Spirit AI believes physical interaction provides information that simulations cannot always reproduce.

"Simulators handle rigid bodies well, but flexible objects like deformable electric cables remain a problem," Gao said.

That creates a costly data problem for the industry. At some Chinese robot-training facilities, operators may have to repeat the same movement more than 50 times to produce one sufficiently precise "clean" example.

Spirit AI has taken a different approach by using what Gao calls "dirty data," consisting of a wider variety of imperfect human movements. The company found that exposing its models to more diverse motions allowed them to improve more quickly, Gao said. The approach reflects a broader challenge in embodied AI: robots need to learn not only how an ideal movement looks, but how physical actions vary when performed by different people and under different circumstances.

From Factory Floors to Homes

Spirit AI's development path highlights why the commercialization of humanoid robots may occur in stages. Factories offer controlled environments where robots can be assigned specific tasks and operate around predictable equipment. Commercial settings such as warehouses, retail locations, and service businesses introduce more variability but can still be designed around defined workflows.

Homes are considerably less predictable.

A domestic robot would need to understand natural-language instructions, identify unfamiliar objects, manipulate items with varying shapes and textures, navigate changing environments, and respond safely around people, children, and pets. That makes household deployment a substantially harder technical problem than demonstrating a robot performing a predetermined movement.

Safety will become another consideration as robots move beyond industrial environments.

The discussion comes as US AI companies face growing scrutiny over autonomous AI agents following incidents involving systems that operated outside intended boundaries. Gao said the immediate risk of a rogue AI controlling a physical robot is lower because current robotic software remains relatively immature.

But that risk could change as embodied AI becomes more capable and robots begin operating around people in commercial and residential environments.

Spirit AI has incorporated physical safeguards into its current systems.

"Our robots feature whole-body force control. If the robot encounters excessive interaction force with the environment, emergency braking triggers automatically as a baseline safety policy," Gao said.

The approach provides a physical layer of protection even when the underlying AI makes an incorrect decision. Gao expects the need for more sophisticated AI safety research to grow as the underlying models become more autonomous.

"Once foundation models reach a mature, autonomous 'GPT-4.0' era, researching advanced AI safety and alignment will become much more actionable," he said.

The trajectory has been touted as an indication that the next major competition in humanoid robotics may be determined less by whether machines can perform impressive physical stunts and more by whether their AI systems can reliably translate language into useful, safe, and adaptable physical work.

China's robotics companies have made rapid progress on the hardware side. The more difficult test now is building the "brain" that can turn those machines into general-purpose workers. If Spirit AI's timeline proves accurate, the first meaningful breakthrough could emerge in industrial settings within the next two years, while the much larger consumer opportunity inside homes may remain further away.