The speed at which global capital is pouring into the embodied AI track remains frantic. With just over nine months passed this year, robotics and Physical AI companies have already secured nearly USD 48 billion (approx. CNY 321.3 billion) in funding, surpassing the total for the entire previous year.
What does this mean in practical terms? Equivalent to approximately CNY 340 billion, an average of over USD 160 million (approx. CNY 1.07 billion) flows into this field every single day.
Billion-dollar funding rounds are emerging one after another, continuously refreshing the valuations of robotics startups, with investment institutions such as NVIDIA, SoftBank, and Sequoia aggressively increasing their bets.
Amidst the boom, however, one cannot help but wonder: where exactly is all this money going? Hundreds of billions of dollars poured in cannot just be for making a splash, can they?
Looking at this year's funding list, it becomes evident that capital deployment has already covered multiple segments of the robotics industry. From manufacturing robot bodies and developing general-purpose intelligent systems, to building training facilities and exploring practical applications, an increasing number of companies are getting involved.
An industrial chain centered around Physical AI is rapidly taking shape.
Global Robotics Funding Hits Record Highs: Which Companies Are Receiving the Money?
Citing the latest data from PitchBook, the Financial Times reports that as of October 8, 2026, global robotics and Physical AI (Robotics & Physical AI) companies have accumulated nearly USD 48 billion in funding this year.
Extending the timeline, from 2022 to 2024, the annual funding for related companies basically remained in the USD 10 billion to 15 billion range; in 2025, the funding scale climbed significantly, and this year it has further accelerated, with the year-to-date funding reaching approximately 1.7 times the total for the entire year of 2025.
Looking at specific funding rounds, leading companies are capturing a substantial portion of the capital.
For instance, the US Physical AI company Anduril closed a USD 5 billion funding round in May this year, reaching a valuation of USD 61 billion, doubling from its previous round.
Anduril is a representative enterprise of Physical AI in the defense sector. This round was led by Thrive Capital and Andreessen Horowitz, making it one of the most watched large-scale funding rounds in this track this year.
Meanwhile, robotics companies targeting scenarios such as industrial and household applications are also continuously receiving capital support.
NEURA Robotics, a German robotics company developing both robots and intelligent platforms, announced a Series C funding round of up to USD 1.4 billion in June this year, with participating institutions including Qualcomm, Amazon, and NVIDIA, among others.
If these companies are advancing robot hardware and intelligent systems simultaneously, then another batch of companies with astonishing funding scales is focusing on the brains of robots.
For example, in January this year, the US robotics foundation model company Skild AI completed a USD 1.4 billion funding round, led by SoftBank, with participation from investment arms under NVIDIA and Bezos, among others, valuing the company at over USD 14 billion.
Another US enterprise developing general-purpose robotics foundation models, Generalist, also announced USD 400 million in new funding in June this year, bringing its cumulative funding to over USD 500 million. Investment arms under NVIDIA and Bezos continued to participate, and AI scientist Fei-Fei Li also appeared on the list of new individual investors.
There is also Physical Intelligence (π), which secured USD 1 billion in funding in the second quarter of this year, continuing to advance its general-purpose models for different robot bodies.
These companies have different technical focuses, but they are all exploring a common direction: freeing robots from their high dependence on preset programs and fixed tasks, enabling them to adapt to more real-world scenarios through learning.
Capital's enthusiasm for this direction is also reflected in the funding structure.
According to a report released by PitchBook in September this year, in the second quarter of 2026 alone, robotics and Physical AI companies raised USD 18.6 billion, representing a quarter-on-quarter growth of 25.3%. However, the number of funding transactions during the same period decreased by 14.1%.
Among these, the three funding rounds for Anduril, NEURA Robotics, and Physical Intelligence totaled USD 7.4 billion, accounting for 39.8% of the total funding for the quarter.
Even excluding these three large funding rounds, the remaining funding amount still reached USD 11.2 billion, a quarter-on-quarter growth of 9.9%.
Therefore, the current funding trend indicates a general trajectory:
Capital continues to pour into the robotics industry, but funds are increasingly concentrated in leading companies and technological directions with platform potential.
Meanwhile, the industrial division of labor presented by the funding list has become increasingly clear.
Some companies are responsible for manufacturing robots and improving hardware performance; some are attempting to develop general-purpose intelligent systems adaptable to different bodies; and others are bringing robots into factories, warehouses, and homes to explore scaled deployment.
Capital is deploying along the entire chain of robotics, from R&D and manufacturing to practical applications.
However, to truly make this chain work, a fundamental question remains.
How exactly should robots learn to do the work?
Tens of Billions of Dollars Pouring In, Robots Are Spending Money to "Go to School" and "Hit the Gym"
This question might sound a bit strange.
After all, large models are now smart enough to handle complex language tasks, and even write code, reason, and plan. So why is it still so difficult to get robots to perform actions like picking up a cup, folding clothes, or tidying a desk?
One of the key reasons is insufficient training data.
For large language models, a massive amount of text, images, and videos has been accumulated on the internet, which developers can directly use for large-scale training. But the data required for robots is much more complex.
For a robot to pick up a cup on a table, it must not only accurately identify the target but also judge its position, size, and shape, plan the motion trajectory, and control the grasping force; change the environment or the task, and it may need to be readjusted.
This information involving space, motion, contact, and force is difficult to obtain completely and directly from existing internet data.
To solve this problem, robotics companies have started trying various approaches.
Perhaps the most interesting of these is the Index program announced by the US humanoid robot company Figure in August this year—paying ordinary people around the world to help collect data for robots.
Users record their daily tasks through the Index app, such as folding clothes, cleaning, and organizing shelves; Figure also sends dedicated recording equipment, and participants record according to the rules and get paid.
After being filtered and processed, these videos can become training materials for robots to learn human actions.
As of August 25, the Index app had accumulated over 264,000 downloads, covering 108 countries and regions, and had collected over 16 million video clips, paying a total of USD 15 million to data contributors.
At the data collection speed at that time, the platform received 30 minutes of video per second, equivalent to adding about 4.9 years of human activity records every day.
Figure also plans to invest over USD 1 billion in data and computing power over the next 12 months. Yes, spending USD 1 billion to learn how humans do the work.
It is worth mentioning that Figure initially tried to purchase training data from third parties, but external suppliers could not meet the requirements in terms of scale, richness, and quality, which led them to shift towards building their own collection system.
Similar explorations are already underway among robotics companies globally.
Physical Intelligence sent robots into multiple Airbnb rentals in San Francisco to test their performance in unfamiliar environments; companies like Generalist, NEURA, and Skild AI are collecting data through self-built training facilities.
The Financial Times refers to such facilities as Robot Gyms—where operators manipulate dual-arm robots to repeatedly perform carrying, grasping, and organizing tasks, while sensors on the robots synchronously record the entire process for model training.
China is also building similar facilities on a large scale. According to statistics from Interact Analysis, as of the end of April this year, there are at least 90 humanoid robot data collection and training centers in China that are in operation, under construction, or in planning, of which 64 have been officially put into use.
As data demand continues to increase, new participants are also emerging.
Companies such as Scale AI and Encord, which previously provided data services for large model enterprises like OpenAI and Anthropic, are expanding their businesses into the robotics field.
Just on October 7, the US startup Mecka, which specializes in providing real-world training data, announced that it has just secured a USD 60 million Series B funding round, led by Sequoia Capital, with participation from NVIDIA, Microsoft's M12, Qualcomm Ventures, Samsung, and others.
Robotics companies are competing to build their own training facilities, data companies that originally served large models are starting to transform, and new enterprises are also receiving capital support.
Thus, it is evident that robot training data is forming an independent industrial segment.
But does collecting more data necessarily mean robots become smarter?
This is precisely the question the industry needs to verify next.
Can Robots Learn to Work Relying on Massive Data?
The willingness of robotics companies to invest huge amounts of money in data is backed by a technical logic that has been verified in the field of large language models—Scaling Laws.
The rapid progress of language models like ChatGPT has already proven the value of large-scale training. Now the robotics industry also hopes to find similar laws: enabling models to master manipulation skills and adapt to new tasks through rich real-world data.
In August this year, research published by the US robotics company Dyna Robotics provided experimental evidence. Its latest world-action model, Dyna-2, was pre-trained using over 1 million hours of human first-person perspective videos, covering daily actions such as cooking, organizing items, and assembling parts.
The researchers set up different training scales of 1,000 hours, 10,000 hours, 100,000 hours, and 1 million hours for comparison.
The results showed that across 14 real-world robot manipulation tasks, as the pre-training data expanded, the average performance of the model improved from 20% to 53%.
The task of turning a key was particularly evident: models trained with less than 100,000 hours of data all failed, while the model pre-trained with 1 million hours of video achieved a 90% success rate. With only about 10 minutes of fine-tuning using robot manipulation data, the pre-trained model acquired the ability to dexterously unscrew bottle caps with both hands.
This research preliminarily proves that large-scale human video data can indeed continuously improve robot manipulation capabilities and increase learning efficiency for new tasks. Although the law needs to be tested in more real-world scenarios, it at least provides a technical basis for the industry's heavy investment in data collection.
However, the quantity of videos is just one factor; data quality, task coverage, and cross-robot transferability equally affect training outcomes.
Therefore, in addition to expanding the scale of data collection, the industry is also exploring another path: bringing robots into real working environments to continuously accumulate experience through actual deployment.
In April this year, Skild AI announced the acquisition of the robotics business under Zebra Technologies (including the former Fetch Robotics).
This acquisition allowed Skild AI to obtain robot products and software systems already applied in warehousing and logistics scenarios.
Compared to relying entirely on laboratory data collection, Skild AI's goal is to leverage existing commercial deployments to bring general-purpose intelligent systems into real warehouses, and continuously improve models using data generated during the actual work of robots.
According to the operational data published by Skild AI last month, about 10 months after its first commercial deployment, its robotic intelligent system has already secured over 60 paying customers, with Annual Recurring Revenue (ARR) exceeding USD 100 million, covering fields such as logistics, manufacturing, security, and food processing.
The company is also collaborating with NVIDIA and Foxconn to deploy Skild Brain on dual-arm robots, exploring high-precision assembly scenarios.
From building dedicated training facilities to letting robots accumulate data in actual tasks, the industry is attempting to establish a more sustainable learning mechanism.
Similar trends can also be observed in the Chinese market.
At the beginning of this year, Liangziwei Think Tank released the "2025 Panoramic Report on Investment and Financing of Embodied AI Startups".
The report shows that in 2025, domestic embodied AI startups experienced a total of 447 funding events, with a total amount of CNY 55.4 billion, while in 2024 it was only 173 events and CNY 13.7 billion. (Note: The report only counts public financing of unlisted companies, which differs from the PitchBook caliber.)
More noteworthy than the scale of funding is that the flow of money has changed.
Liangziwei Think Tank found that in 2024, domestic embodied AI financing was still concentrated in robot body enterprises. By 2025, as leading companies entered the stage of mass production and implementation, funds began to diffuse to industrial chain segments such as embodied brains, core components, training data, and vertical applications.
Among them, Yinhe Tongyong refreshed the record with a single funding round of CNY 2.031 billion; the synthetic data company Guanglun Intelligent also completed a funding round of several hundred million yuan, with revenue breaking CNY 100 million.
Meanwhile, domestic internet giants, automakers, manufacturing enterprises, and state-owned funds are also increasing their deployments.
These industrial capitals bring not only funds but also supply chains, production lines, and practical application scenarios. In 2025, strategic investments have become increasingly active, with capital helping robotics companies acquire landing opportunities through investments.
This is particularly meaningful for the Chinese market, which possesses a massive manufacturing system.
Real production environments such as factories and warehouses can not only verify robot capabilities but also continuously generate new training data—the experience accumulated by robots in different tasks, after processing, can feed back into the models to support more complex work.
Currently, robotics companies are still spending heavily to purchase equipment, hire people, and build training facilities, specifically to produce data for models.
However, as robots enter actual production environments, commercial deployment itself also has the opportunity to become a continuous source of data. More deployments bring richer data, and the improvement in model capabilities creates conditions for robots to enter new scenarios, thus extending the scale effect.
Global capital is betting on foundation models, while Chinese capital is perfecting the industrial chain; behind this is a similar direction: robot model enterprises proactively acquire commercial deployment resources, while industrial capital brings robots into factories and production lines.
The value of real-world application scenarios is extending from verifying robot capabilities to supporting robot learning and evolution.
Spending money to hire people to shoot videos and operate robotic arms... is ultimately phased. What is truly worth looking forward to is robots starting to accumulate their own experience and become smarter on their own in real work.
This is perhaps the more exciting industrial change after tens of billions of dollars are poured in.
— End —