The previous article “Only 8 months left! AI brain undergoes a complete overhaul” calculated the account clearly: WAM optimizes the single step action probability P (a ₜ | o ₜ, I), while WTM optimizes the task success rate P (success | τ) of the entire trajectory.
The 10 step task has 95% of each step, leaving only 60% for multiplication – this is mathematics, not viewpoint.
Since WAM has already reached LIBERO 99.25% and RoboTwin 94.3% in benchmark testing, which seems “sufficient” in a single step, why can’t the industry wait for WAM to grow into WTM on its own? Why predict the urgent time window of ‘within 8 months’?
The answer is not in the paper, it is in the account book of Party A. In 2026, there have been 42 financing events in the world model track, basically equaling the entire year of 2025, with a total financing amount of over 4 billion yuan in the first half of the year.
Money is pouring in crazily, but the assessment criteria for funds have completely changed – WAM’s “one-step myth” is colliding with the “global KPI” of the production line.
WAM’s’ Technology Self Improvement ‘and Party A’s’ Survival Metrics’
The report released by KPMG during WAIC is very straightforward: the investment evaluation dimension has completely departed from the old framework of “comparing model parameter quantity and demonstration effect” in the past two years, and has fully shifted to three engineering indicators: task completion rate in real scenarios, generalization ability in unfamiliar environments, and unit deployment cost. Project due diligence is shifting from “reading papers and presentations” to “looking at production lines and order repurchases”.
KPMG China’s Lu Lingpeng further clarified the criteria for the “commercialization singularity”: only when the model “accurately commands robots to complete long process tasks within a very short decision-making cycle, significantly improves the rate of manual substitution, enhances equipment operation efficiency, and accelerates production pace”, can it truly cross the singularity of commercialization landing.
Pay attention to these three key words – long process tasks, equipment operation efficiency, and production pace. There is no single step success rate.
This is precisely WAM’s Achilles’ heel. The WAM manufacturer presented “single step 95%”, “LIBERO 99%”, and “inference 85 milliseconds” during the bidding process, but the first party wrote a different set of words in the bidding documents:
Task completion: Can the yield of this line be improved by 5 points?
Conversion efficiency: How long do I have to stop production for changing SKUs? How much pressure can you achieve in a week using traditional methods?
Unit deployment cost: How much labor can this machine save me and how much scrap can it reduce in a year?
Return cycle: When can I retrieve this CAPEX?
In the bidding documents of the top 3C factories, “task completion rate” and “changeover efficiency” have been listed as veto items. Using the ‘95% success rate per step’ to deceive, even the bidding threshold cannot be reached.
Why is Party A suddenly so ruthless? Because in the first half of 2026, the total financing amount for the domestic embodied intelligence track exceeded 90 billion yuan, a fivefold increase compared to the same period last year. When the money is burned to this level, the demo period must end. DaXiao Robot has successfully implemented a complete closed-loop system of “scene collection data iteration model optimization revenue return” in retail scenarios such as PetroChina convenience stores and Kuaikeda, and has officially announced the completion of a thousand store landing within the next year. This means that some people in the industry have started to use the term “cost recovery”, and everyone will be measured by this standard.
Three unsettled accounts of WAM on the production line
Aligning the technical features of WAM with the KPIs of Party A one by one, it will be found that there are three WAM accounts that cannot be calculated clearly, which is also the biggest obstacle to the commercialization of WAM.
First account: Accumulation of long-term errors. KPMG clearly listed “the accumulation of long time series errors reduces the reliability of virtual and real migration” as one of the three core difficulties in the commercialization of the world model: the law of model learning in the virtual environment. Once there is a small deviation in the actual implementation, all subsequent predictions will deviate from the real track, and the longer the task, the greater the deviation. WAM optimization focuses on “whether the next action is correct”, and it does not have a “task level correction” mechanism in its architecture – errors will not be absorbed in the task progress space, but will only multiply all the way to the end of the task. This is WAM’s original sin, not a bug.
Second account: Replacement cost. The mainstream VLA/WAM route relies on massive pre training data, and the pre training data becomes completely ineffective when the object surface is wet or deformed. The first party changing to a new SKU means re collecting, re labeling, and re adjusting. The traditional automation solution takes more than a week to switch production, and WAM manufacturers have not fundamentally shortened this time – because their “world” is modeled during training, and they have turned off “looking to the future” for 85 milliseconds during inference. The first party requires a ‘minute level production change’, while WAM provides a ‘weekly level fine-tuning’, which is a difference in magnitude.
The third account: real-time contact operation. The contact operation in industrial scenarios requires real-time feedback at the hundreds of Hz level, and the WAM large model scheme is difficult to achieve such a fast and immediate response due to inference latency. What Party A wants is “the hose should not slip, the bottle cap should not overflow, and the assembly should not get stuck”, while WAM gives “it looks elegant” – another misalignment.
The commonality among these three accounts is that each technical indicator of WAM is beautiful, but each indicator does not directly correspond to the KPI of Party A.
WAM optimizes the “elegance of actions”, while Party A pays for “reliable completion of tasks”. This is a fundamental misalignment of optimization goals, not a gap in engineering details.
Capital has voted with its feet: from ‘model worship’ to ‘landing is king’
The more deadly signal comes from the financing side. On August 10th, Oakfruit Robotics completed an angel round financing of hundreds of millions of yuan, jointly led by China Merchants Capital and NIO Capital, with follow-up investment from Shuimu Tsinghua Alumni Seed Fund. It has only been four months since its nearly 100 million yuan seed round completed in March. The choices of capital in these two industries are very representative – China Merchants Venture Capital focuses on ports and industrial manufacturing, while NIO Capital focuses on high-end manufacturing and intelligent hardware. They are not investing in ‘world models’, they are investing in’ things that can work in factories’.
The commercial data of oak fruit Natus AGE-0 can better illustrate the problem:
It took about two months from signing the contract to passing the POC verification on the ODM production line of the world’s top cosmetics manufacturers; The time for production and debugging has been reduced from several weeks in traditional solutions to 6-40 minutes; Targeting discrete manufacturing scenarios with multiple SKUs, small batch sizes, and frequent production changes in fast-moving consumer goods, food, 3C, and other industries; We have established cooperation with multiple global top customers from different industries to complete POC verification of the production line for grasping, assembly, twisting, and plugging.
Why did China Merchants Group and NIO invest in oak fruit? Because they can understand what “6-40 minute production change” means for Party A – this means that the production line can handle orders of multiple varieties and small batches, which means an increase in output per unit time, and that this CAPEX can really recoup costs.
On the other hand, WAM manufacturer X-Era Lab is doing “symmetrical stitching of plush bear ears” at a toy factory in Dongguan. The single machine has an average of 187 grabs per day, a success rate of 98.4%, and a beat time of 22.6 seconds – the data is very beautiful, but this is the ultimate optimization of a single SKU. What about changing to a plush toy style? The same is true for CR Vanguard replenishment robot. The current commercialization progress of WAM manufacturers is mainly focused on “extreme optimization of single SKU” or “relatively standard scenarios such as retail/warehousing”. Once they enter the true discrete manufacturing of “frequent production changes and complex materials”, the single step optimization advantage of WAM will be swallowed up by the cost of production changes.
Capital has seen through this.
So KPMG would say that ‘the investment evaluation dimension has completely departed from the old framework of comparing model parameters and demonstrating effectiveness’; So Xia Zhijin, the managing partner of Xiangfeng Investment, will publicly criticize the valuation of billions of yuan as an “emotional expression, stemming from unrealistic expectations of technological progress”.
The true meaning of the 8-month countdown is not that the technology is mature, but that the first party has exhausted their patience
Now we can answer ‘why 8 months’. This is not a cycle of technological maturity, but the limit of patience of Party A, the pressure of capital recovery, and the time point when WAM’s three unsettled accounts on the production line are due at the same time:
The requirements for real-life training in 2026 issued by the Ministry of Industry and Information Technology and the State owned Assets Supervision and Administration Commission require that humanoid robots must achieve normalized effects on production lines before the end of the year;
Top 3C factory bidding documents: task completion and changeover efficiency are rejected with one vote;
The core indicators for KPMG to determine the singularity of commercialization are: success rate of long process tasks, production pace, and equipment operation efficiency;
Capital payback pressure: Financing of 90 billion yuan in the first half of 2026, five times year-on-year, money cannot be burned forever on demos;
These four time points overlap to form the final deadline for WAM manufacturers to provide a “task level answer” – starting from August 2026, which happens to be 8 months, the scale landing acceptance window for Q1-Q2 in 2027.
What will happen at that window?
Either PI and Xingyuan Intelligence explicitly add the “task progress regression head” to the joint training objective in the shell of VLA/WAM, allowing WAM architecture to evolve into WTM; Either Kuwa can transition WAIM from continuous driving to discrete operation (with a significant architecture gap); Either acorns, as an “outlier,” rely on Natus AGE-0’s millisecond level instinctive reflexes and minute level production changes to lead the way in the “plug and play” commercial loop on the production line, forcing the industry to acknowledge that “WTM does not necessarily have to exist in the form of a world model”; Either the pure engineering WAM, such as the Excellent Vision, is forced to turn back and retrieve the ‘world’, along with the task map – the cost of transformation is not small.
But no matter how it evolves, WAM, as an independent optimization paradigm, will retire within 8 months. Its technological heritage will be absorbed by WTM (sub tasks/subgoal of PI, Envision Verify Act closed-loop of StarSource Intelligence, WAIM recursive correction of Kuwa), but its first principles of “single step action optimization” will be abandoned.
相关文章
