Gasgoo Munich- Two years ago, the biggest fear for investors in embodied AI was missing out.
As soon as a project launched, investors scrambled to get in, driving valuations higher with every round. Everyone feared that a moment's hesitation would mean losing their seat. Back then, the race was about who moved fastest and who dared to bet the biggest.
Now, the fear is stepping on a landmine.
Where the money goes, how much is handed out, and the valuation attached to it have all shifted compared to six months ago.
Deal Volume Remains High, but Mega-Rounds Hit the Brakes
By the numbers, financing in the embodied AI sector remained elevated in October.
According to incomplete statistics from Gasgoo Embodied AI, 49 domestic companies in the robotics and core components space disclosed new funding or investment updates that month, sustaining the momentum seen in recent months.
Yet, behind that high volume lies a pullback in mega-deals and a shift in investment strategy.
On one hand, investors are still betting early and small. Of the 49 funded companies in October, 16 disclosed seed or angel rounds — roughly a third of the total.
On the other hand, large-ticket funding rounds declined markedly. Only four companies disclosed deals in the RMB 1 billion range during the month, namely Astribot, FAIRINO, Huixi Technology and MOSS Intelligence. For both Huixi and FAIRINO, the figures are combined totals across multiple rounds.
Image Source: Astribot
By contrast, across all segments in the first half of the year, 38 funding deals were in the RMB 1 billion range, with April and June each seeing as many as ten. In April, TARS disclosed a Pre-A round worth as much as USD 455 million, setting two records at once in China's embodied AI sector: the largest single funding round and the largest Pre-A round on record. In June, AI² Robotics announced a string of new funding rounds with a combined total of close to RMB 5 billion.
Today, while the flow of deal announcements remains active, the number of 1 billion yuan projects has dropped significantly compared to the first half.
Viewed together, the trend is clear: the market has not sustained its earlier breakneck expansion. Investors are applying stricter filters to screen projects.
As standards shift, so does the destination of capital.
In the first half, heavy capital flowed into key components like general-purpose hardware, "brains," data, and dexterous hands. Entering the second half, upstream sectors — tactile sensors, joints, reducers, encoders, and thermal management — began securing funding more frequently. Overall, the upstream "shovel sellers" remain a stable harbor for capital.
At the same time, a cohort of companies focused on vertical scenarios has entered the spotlight.
The makeup of investors is shifting, too. In the funding rounds and shareholder changes disclosed in October, robotics manufacturers appeared frequently.
AgiBot, for one, has spread its bets across multiple segments at once. It appears on the new shareholder rosters of embodied-data firm Yitai Jushen and robot thermal-management maker Wenqing Zhikong, and it also backs space-robotics company SINXBOT. Back in February, AgiBot joined the Pre-B funding round of Huixi Technology.
X Square Robot, meanwhile, took part in the latest funding rounds of integrated-joint maker Juxie Intelligent Drive and tactile-sensor company Matrix Innovation. Fourier and Galbot co-invested in powertrain firm YITA Dynamics, while LimX Dynamics showed up among the co-investors in full-body tactile sensing startup MoSense.
Image Source: Unitree
Also in late October, Unitree announced that it had jointly set up the "Shanyu Chuang" initiative with HongShan. Dedicated to sourcing early-stage projects, the program targets embodied AI, robotics applications and key upstream and downstream technologies.
If previous single investments were merely deal-by-deal, this partnership extends robotics manufacturers' reach from one-off bets to a platformized, institutionalized hunt for early-stage opportunities.
Why are robotics manufacturers moving aggressively into the supply chain now?
A key reason is that the battleground has shifted. Two years ago, the race was about demos and specs. Today, supply chain stability and cost control are determining a company's ability to deliver. By positioning themselves here, manufacturers are not just bolstering their supply chains and technology stacks — they are improving coordination efficiency.
Yet, this tight coupling carries risks. If collaboration falls short, valuations fluctuate, or the manufacturers' own production volume lags, those risks could ripple through equity ties to the broader ecosystem.
But at least in October, top-tier manufacturers put real money behind their stance. The competition in embodied AI is expanding from solo breakthroughs by individual firms to a contest between supply chains and ecosystems.
Vertical Scenarios Become a Key Destination for Capital
Looking back at the early days of autonomous driving, after the Robotaxi hype cycle, the industry gradually fragmented into specific scenarios. Embodied AI is now undergoing a similar evolution.
The October data is blunt: 19 companies that disclosed funding were focused primarily on vertical application scenarios — nearly 40% of the total.
Where is the money going?
Based on the rationale for deploying robots, these projects fall into three broad categories.
The first serves daily life.
For example, Dabai Robotics focuses on elderly care and companionship, Superpi on mobility for seniors, Zhuge Robotics on guidance for the visually impaired, Ola Dimensions on home scenarios, and Jingji Intelligence on building cleaning, while PongBot brings an AI coach right to the table-tennis table.
The second category enters production segments where precision and stability are critical, and labor costs are climbing.
In healthcare, Shendu Medical, abrobo and Deyeemed specialize in microsurgery, vascular intervention and TCM physiotherapy, respectively. On the factory floor, Niuer Robotics focuses on semiconductor packaging, SRT on automotive welding, CASBOT on underground mining and shoemaking, and Tsing-AI on the intelligent retrofitting of construction machinery.
The third category ventures into places humans cannot reach — or dare not stay for long.
Image Source: SpaceRobot
Consider the space-facing players such as SINXBOT and SpaceRobot, or the underwater specialists CHASING and Weilan Deep Sea. For high-risk posts such as refueling and firefighting, Ruoyu Tech and Zhongke Bote are also in the field.
Notably, judged by their founding dates, these companies include both new teams set up in the past couple of years and long-established firms. Zhuge Robotics, founded back in 2009, has spent 17 years working in accessibility for people with disabilities and in guide robots.
Why have vertical scenarios begun to draw such broad attention from capital?
To answer that, one first has to see where the industry was previously stuck.
Over the past two years, the valuations of many general-purpose humanoid companies ran ahead of their deliveries. Demonstrations kept multiplying, yet only a limited number of firms could prove continuous operation, scaled delivery, and profitability.
After entering 2025, market divergence consequently widened. One pool of capital kept betting on general-purpose capabilities, reaching a peak in the first half of this year; another raised its requirements on orders, delivery, and payback periods.
On the whole, capital has not fully withdrawn from the general-purpose track, but demonstrations and long-term vision alone are no longer enough to support every company's valuation.
The choice made by GSR Ventures offers a useful sample.
In March 2025, Allen Zhu, managing partner of GSR Ventures, said publicly that the firm was exiting a batch of humanoid-robot companies. By June 2026, GSR had doubled down across five consecutive rounds in marine-robotics company SEAHI Robotics.
In a speech at Peking University this September, Zhu went further with a rhetorical question: "China now has more than 200 companies making humanoid robots—which one can't dance? Which one can't throw a punch?"
His judgment was direct: a large number of robot makers offer similar capabilities, yet commercial demand is insufficient to accommodate so many companies.
One institution's choice may not represent the entire industry, but public statements from several investors and the structure of the deals show that project evaluation now increasingly factors in repurchase rates, continuous operating hours, real orders, and delivery capability.
That said, a shift from general-purpose to vertical does not mean every vertical can make the numbers add up. A line has to be drawn first.
The scenarios capital can assess more readily are those that already come with orders, customers, and deployment channels, and whose value can plausibly be verified by workload, yield, or cost reduction.
Image Source: CASBOT
For example, LeadMind disclosed that its orders on hand plus intended orders for 2026 approach RMB 1 billion, with customers including Micron and SK hynix, of which cumulative orders from Micron alone amount to several hundred million yuan. CASBOT's 2026 backlog for embodied shoemaking robots has also surpassed RMB 400 million, with scheduled production moving into the thousands-of-units range. Deyeemed has cumulatively served more than 3,000 industrial partners, and Chlitina's store network provides it with a potential deployment channel.
For these scenarios to generate sustained procurement, everything ultimately comes back to concrete metrics: hospitals focus on stability, safety, and clinical value, while factories look at yield, production takt time, labor costs, and downtime losses.
Conversely, in scenarios such as elderly companionship and guidance for the visually impaired, although demand is real, the mechanism for large-scale, sustained payment remains unclear. Who pays for companionship—families, the government, or insurers? Busy as these scenarios look, the economics are still hard to work out.
Moreover, these verticals are not all blank-slate markets. In shoemaking, welding, and packaging, companies such as FANUC and ABB have long had a presence. In other words, newcomers must both compete with established automation vendors and persuade customers that still rely on manual labor to rework their processes.
In that sense, narrow scenarios may generate revenue faster, but they are not necessarily easy to grow into large businesses. Capex swings within a single industry, large-customer concentration, acceptance cycles, and accounts-receivable problems all affect whether orders actually convert.
An even more direct constraint is replicability—whether a solution can expand from one factory to ten, or extend from one process to adjacent ones. If every new contract requires fresh customization, revenue growth will simultaneously push up spending on engineers and after-sales support.
Overall, then, the attention on vertical scenarios reflects only that capital has raised its bar on the speed of realization. Orders are merely the first token of proof; delivery, acceptance, payment collection, and repeat purchases still lie behind it. Only companies that can work through these steps and then replicate across customers are genuinely closer to a real business.
Physical RSI: Betting on a more distant future
If investing in concrete scenarios is about finding a landing point for near-term realization, the rising enthusiasm for Physical RSI is a bet one step further out.
RSI stands for Recursive Self-Improvement, meaning that AI takes part in improving—and even building—the next generation of AI, forming an "AI improves AI" loop.
The goal of Physical RSI is to place this loop in the physical world: a robot spots its own shortcomings while working, generates new tasks and training data, updates its policies, and then takes the new version back into reality for validation, with the results triggering the next round of improvement.
In this cycle, each "new version of itself" that is output becomes the input for the next round, producing a recursive improvement process.
Physical RSI and kindred self-evolution directions have recently drawn wide attention. From leading humanoid makers such as Unitree and AgiBot to a host of startups, all are actively positioning themselves.
Back in August, Unitree founder and chairman Wang Xingxing revealed at the World Robot Conference that the company is pre-researching a self-evolution system for physical-AI robots, letting AI take over R&D to achieve self-evolution of physical-AI robot models.
Specifically, using cutting-edge foundation models as the core, Unitree first sets the boundaries of skills, tools, and tasks; the model then searches the web on its own for the latest papers, strong research, and open-source solutions, builds its own plans, and generates robot control code, which it runs in simulation or hands to other large models before deploying on real hardware for testing. Finally, AI and humans score the results together and feed them back to close the loop.
"This entire chain can perfectly be automated by AI: AI writes the code, trains, deploys, and runs automatic tests itself; AI or humans score it; and then it iterates on its own," Wang said.
Image Source: AgiBot
AgiBot, meanwhile, released the embodied-agent framework Genie Master and the full-chain learning model GO-3 Preview back-to-back in October, providing the technical foundation for Physical RSI and proposing two key mechanisms, Harness RSI and Policy RSI.
Harness RSI mainly keeps a continuous log of successes and failures and summarizes them by day to form reusable long-term memory that improves task planning. Policy RSI sends the execution and recovery trajectories from real operations back into training, directly updating the parameters of motion policies.
On November 2, AgiBot put six AgiBot Genie G2 humanoid robots into a convenience store in Xuhui, Shanghai, to "work" around the clock for 24 hours. AgiBot says this is the industry's first validation of Physical RSI in a real business scenario.
Image Source: NEXFORM
Simate and NEXFORM have also laid out explicit Physical RSI plans; Ola Dimensions calls its technology system Real-World RSI, while Gamma Robotics emphasizes continuous learning through real-world interaction and self-evolution of full-body models.
All of these companies have recently closed new funding rounds. Simate simultaneously won investment from both Xiaomi and Baidu; Gamma Robotics was incubated by China Telecom's AI research institute TeleAI; and Ola Dimensions has raised four rounds in succession within just half a year of being founded.
However, although these directions all involve self-evolution or continuous learning, the industry currently has no unified evolution path or grading standard.
A common approach is to divide RSI into weak, medium, and strong stages according to the role humans play.
In weak RSI, humans define the goals, boundaries, and acceptance criteria, while AI mainly runs experiments, tunes parameters, and makes initial selections within a fixed framework, serving as an R&D assistant. In medium RSI, the AI can identify some problems on its own, propose and test improvement hypotheses, with humans intervening only at key decisions. In strong RSI, the system must be able to improve the "improvement mechanism itself," allowing R&D capability to self-accelerate generation after generation.
A finer five-layer division also exists.
Chaojinhua, for instance, splits RSI into five layers from shallow to deep. The shallowest is the output layer, where automated research lets research outputs self-correct; next is the experience layer, which continuously updates runtime assets such as memory and skills; then the Harness layer, where the agent's own code keeps iterating; the fourth is the model layer, which truly internalizes new capabilities through continuous parameter updates; and the deepest is the evolution-mechanism layer, which hands the very methodology of "how to evolve" over to AI itself.
Whichever taxonomy is used, public demonstrations suggest most systems still require humans to take part in goal-setting, evaluation, or key decisions; no recognized evidence yet exists of strong RSI running continuously in an open physical environment.
Wang Xingxing has previously stressed that in today's self-evolution systems, humans still have to take part in scoring. Whether a motion is appropriate, whether the success rate is sufficient, and whether the outcome is truly what people want—AI cannot yet judge all of this on its own.
Why is Physical RSI heating up at this moment?
One important driver is that, on some current models and tasks, continuing to expand pretraining data and compute has not brought commensurate gains in real-world performance. Meanwhile, advances in coding agents have made it easier to chain code generation, experiment execution, results analysis, and part of the iteration into automated workflows.
More specifically, three conditions have converged.
First, the digital world has already validated the feasibility of "AI doing R&D."
AI code-writing and AI-assisted research tools keep getting stronger. Anthropic, for example, previously disclosed that as of August 2026, across its internal AI R&D tasks, Claude, under human supervision, was already leading about 26% of the work, though none of the tasks tested had reached fully autonomous operation.
The digital world has already automated parts of R&D—code generation, experiment execution, and results analysis—providing an engineering workflow the physical world can borrow. Simate's AutoResearch is essentially this automated R&D approach carried into robotics.
Second, the engineering pipeline for embodied intelligence is also coming together.
The infrastructure for VLA (vision-language-action) foundation models, world models, high-fidelity simulation, reinforcement learning, and post-training has matured to a certain degree. This maturity has lowered the engineering threshold from model modification and simulation validation to real-hardware feedback, making continuous iteration easier to achieve.
A third, easily overlooked condition is that robots are genuinely starting to "take up posts."
As more robots go to work across scenarios, some companies are beginning to accumulate operational data that can flow back continuously, providing fuller conditions for post-deployment learning—which is especially important for Physical RSI, with its emphasis on learning after deployment.
Overall, however, the current hype around Physical RSI is largely the combined result of technical progress, capital narrative, and academic naming. Although public results span pre-research projects, technical frameworks, and small-scale scenario validation, cross-scenario, long-cycle, reproducible data is still lacking, and the field remains some distance from an engineering closed loop of stable operation and replicable scaling.
Conclusion
Looking back at October, financing in embodied AI is hard to sum up with a simple "hot" or "cold."
Money is still entering the sector, but the tests applied to different projects have diverged: humanoid makers must prove that industrial investment can translate into stable supply and scaled delivery; vertical-scenario companies must prove that orders can turn into delivery, acceptance, and payment collection; and Physical RSI must produce cross-scenario, long-cycle, reproducible improvement results.
Capital has raised its demands for near-term realization, but its interest in longer-dated technology has not disappeared, and technology directions that have yet to complete commercial validation can still secure funding.
Going forward, the track's deeper divergence is expected to unfold after financing. Only those that can turn technology into sustainably deliverable products, and orders into collected payments, will truly have crossed the threshold of commercial validation.
