PRIMACT R100 puts local inference before general factory robots

Zhílái Embodied Intelligence is building a dual-layer system that pairs a general large model with expert models, reporting 99% model accuracy. The report attributes a tougher factory threshold to Yu Qiong: 99.99% is needed before production lines can be connected, while each move from 99% to 99.9%, then to 99.99%, doubles R&D investment.

That gap changes what counts as a commercial robot.

ifanr reports Zhu Xing's view that deployable robots largely rely on imitation learning and are especially exposed to unseen exceptions; even convenience-store deployment at scale remains constrained by capability and cost. Zhílái plans robot products with 3-5 skills simultaneously by mid-2027. The useful synthesis is how bounded work, local inference, expert modules, and recovery from exceptions can turn limited skills into a repeatable factory purchase.

A production line sets a harsher accuracy threshold

Zhílái Embodied Intelligence's stated 99% model accuracy illustrates why a factory line is a tougher first market than a broad demonstration. According to the outlet, Yu Qiong argues that industrial scenarios need 99.99% accuracy before they can connect to production lines. The difference is small on paper. On a line, it defines whether the system can be trusted to keep work moving.

The last fraction of reliability is also expensive.

Yu Qiong told the outlet that moving from 99% to 99.9% requires R&D investment to double, and reaching 99.99% requires it to double again. That creates a commercialization constraint: a robot must earn its way through a narrow, measurable task before a buyer can justify the engineering needed for line-level reliability. A general model can contribute, but it does not erase the cost of handling the cases that remain.

Zhílái's approach reflects that trade-off. Its dual-layer architecture combines a general large model with expert models, and the company says it plans robot products with 3-5 skills simultaneously by mid-2027. This is a more bounded proposition than a machine expected to shift freely among unfamiliar jobs.

ifanr's Zhu Xing makes the same pressure visible from deployment. He says even convenience-store deployment at scale remains difficult because capabilities and costs are inadequate. Current mainstream deployable products rely mainly on imitation learning, he says, leaving them especially exposed when an unseen exception appears. Industrial entry points therefore depend less on a robot's range of advertised skills than on the accuracy and exception tolerance a specific line can afford.

Local inference is part of the factory boundary

PRIMACT R100 frames local inference as part of the factory boundary. According to a report, Yuanshi Intelligent's platform carries core model algorithms alongside industrial interaction interfaces, and supports both local deployment and local inference. That combination matters because a retrofit device must connect intelligence to equipment without assuming that production data can leave the site.

The restriction is organizational before it is computational. Xu Chen told reporters that, after visits to more than 200 companies, the team found factories generally would not network data because of core-confidentiality management requirements.

Local deployment therefore changes what an industrial AI product must contain. Yuanshi's software toolchain is intended to develop and adapt industrial AI models and process packages for particular equipment and process scenarios. The platform is not simply a place to run a model; it is a boundary where model logic, process knowledge, and industrial interfaces can be kept close to the machine. Xu Chen's view, reported in the same account, is that coordinated large and small models are the viable industrial-AI route because computation must be localized.

Real-time behavior makes that boundary tighter. Zhu Xing told ifanr that robot world models must obey physical laws and perform highly in real time, unlike digital video models that can tolerate latency in pursuit of visual quality. In a factory setting, local inference is therefore tied to response timing and confidentiality, not hardware specifications alone.

Reusable modules still need site-specific adaptation

Reusable modules are a way to package hard-won shop-floor knowledge without claiming that every new site is already solved. According to a report, Yuanshi Intelligent is treating actions such as Grasp, Twist, and Insert as physical modules that can be combined for new tasks. Its WMM adds dynamic structures and physical constraints to model learning, aiming to identify patterns from data rather than requiring coverage of every operating condition.

That approach can preserve expertise at a useful level of abstraction. Over two years, Hemu Intelligence organized 100,000-level high-quality drawing data and drew on more than 100 veteran mold workers to build an industry knowledge system, the report says. It has also trained more than 1,600 small models for high-precision execution in different manufacturing scenarios. The module is reusable; the evidence behind it is still highly specific.

Replication changes the evidence requirement.

According to ifanr, Zhu Xing argues that deploying across pharmacies and convenience stores remains costly because models do not generalize sufficiently between stores. Layout changes, equipment differences, and unusual cases can shift the data distribution enough to make a familiar action unreliable. More data does not automatically close that gap: Zhu Xing said that using 3 times or 5 times more data can worsen overall results when the distribution is heavily imbalanced, even if some task types improve.

The commercial discipline is therefore to collect the data that matches the next operating environment. ifanr reports that Ant Lingbo moved early toward customized data collection instead of buying large amounts of finished data, raising the usability rate in its internal data funnel from about 15% to more than 90%. Reusable actions reduce the amount of work that must be rebuilt. They do not remove the work of adapting a process package to a new site.

Humanoid production is not yet the commercial benchmark

XPeng's humanoid effort shows why production capacity and commercial proof should be treated as separate milestones. According to geekpark, the company independently designed and developed what it calls the world's first automated production line for advanced humanoid robots, with core-process automation exceeding 80%. That describes a way to build machines repeatedly. It does not by itself demonstrate a repeatable operating cycle at customer sites, where interruptions and operating economics decide deployment. That distinction matters.

The stated first destination is internal. geekpark reports that XPeng plans initially to place robots in its own stores and industrial parks, before planned large-scale mass production by the end of 2026.

The vehicle path is more established in a specific sense. woshipm says XPeng deployed its second-generation VLA model in vehicles last year and describes it as its first mass-produced physical-world large model, intended for cars, Robotaxis, robots, and flying cars. In May, XPeng's first self-developed mass-production Robotaxi vehicle rolled off a production line in Guangzhou. Those facts support a vehicle deployment route, but they do not establish that humanoids already have a commercially closed loop. ifanr attributes to Zhu Xing the view that even convenience-store deployment at scale remains difficult because capabilities and costs are still inadequate.

How Chinese physical-AI firms turn constrained industrial work into deployable systems

Choose physical AI by constraint, locality, and repeatability

• A factory has a tightly defined, high-precision task such as inspection, grinding, ironing, or a repeatable mold-engineering workflow. Start with a narrow task-specific deployment rather than a general-purpose humanoid. Zhílái Embodied Intelligence says its flash-shot inspection, force-controlled grinding, and ironing robots have entered automotive-parts production lines in batches, while Hemu Intelligence has applied AI to process programming, intelligent quotation, and process-route planning.
• Equipment data or production know-how cannot be connected to external networks because of confidentiality requirements. Favor an edge deployment with local inference and equipment-specific process packages. Yuanshi Intelligent's PRIMACT R100 supports local deployment and local inference; Hemu Intelligence says factories it surveyed generally do not allow data to be networked because of core-confidentiality management requirements.
• The workflow is a combination of recurring physical operations, such as grasping, twisting, and inserting, across related equipment or tasks. Encode those operations as reusable modules before attempting broad generalization. Yuanshi Intelligent is attempting to turn Grasp, Twist, and Insert into reusable physical modules that can be combined and adapted for new tasks, supported by a toolchain for adapting industrial AI models and process packages to specific equipment and process scenarios.
• A deployment must meet production-line reliability requirements, not merely deliver an impressive demonstration. Set the acceptance bar around the line's required accuracy and budget explicitly for the final reliability increments. Zhílái Embodied Intelligence says its general-large-model-and-expert-model architecture achieves 99% model accuracy, but Yu Qiong says industrial scenarios require 99.99% accuracy and that moving from 99% to 99.9%, then to 99.99%, requires successive doublings of R&D investment.
• A team wants to roll the same robot solution out across multiple stores, sites, or changing environments. Prove one constrained site first and treat cross-site replication as a separate engineering problem. Ant Lingbo's Zhu Xing says replication across stores is costly because robot models lack sufficient generalization across different pharmacies and convenience stores; its pharmacy deployment used shelf aisles only 80 centimeters wide.

Procurement should start with recovery, not a world model

Procurement should begin with a failure definition and a quality threshold, rather than a broad claim about intelligence. Yu Qiong told a technology outlet that industrial scenarios need 99.99% accuracy before production lines can be connected. Buyers should ask which task is measured, what counts as an error, and how performance changes when the input differs from the training case.

Recovery matters because failures are visible. In February 2026, IRON lost balance and fell while walking during its first public appearance at Shenzhen Bay MixC, according to geekpark.

He Xiaopeng's view, reported by geekpark, is that robot development requires accepting failure and that growth carries a cost. That makes exception handling a purchasing requirement: can an operator stop the system, return it to a known state, and record what triggered the intervention? Zhu Xing told ifanr that mainstream deployable products rely mainly on imitation learning and are especially exposed to unseen exceptions.

Local-data control should be assessed alongside adaptation effort. Zhu also told ifanr that replication is costly when models do not generalize sufficiently between pharmacies and convenience stores. A buyer should require a clear account of site-specific work, then ask unresolved due-diligence questions: Which safety certification applies? Is there a rollback path after an upgrade? Are audit trails retained? Can humans override decisions? What upgrade path preserves validated behavior?

Set procurement gates around the data loop, not only model scores. Ask how exceptions will be logged, labeled, and returned to training; Ant Lingbo says customized collection lifted usable data in its internal funnel from about 15% to more than 90%, according to ifanr.

Do not assume more footage fixes a weak deployment. Zhu Xing warns that 3 times or 5 times more heavily imbalanced data can worsen overall results, even when selected tasks improve. Require evidence that training data covers each site's variation and that post-training can use few-shot examples efficiently.

Budget explicitly for adaptation between locations. Replication across pharmacies and convenience stores remains costly because current robot models do not generalize sufficiently between stores, Zhu Xing said. Local systems also need real-time performance and physical-law consistency, rather than the latency tolerance of digital video models. Watch for native robot training paths: LingBot-World 2.0 released a 1.3B small model and part of its training technology.

For readers outside China

• Availability: Availability outside China is largely not disclosed in sources. XPeng plans to initially deploy its robots in its own stores and industrial parks, plans to begin large-scale mass production by the end of 2026, and says XPeng IRON is planned to officially go on public sale worldwide in 2027. The source material does not cover overseas sales, support, or deployment availability for Zhílái Embodied Intelligence, Yuanshi Intelligent, Hemu Intelligence, or Ant Lingbo.
• Pricing: Robot and software pricing is not disclosed in sources. The source material reports financing and valuation figures for some companies, but does not state customer purchase prices, subscription prices, or deployment costs.
• Closest Western equivalents: The source material does not identify Western equivalents.
• Data residency: The strongest documented pattern is local industrial deployment rather than a stated jurisdictional data-residency policy. Yuanshi Intelligent's PRIMACT R100 supports local deployment and local inference. Hemu Intelligence says factories generally do not allow data to be networked because of core-confidentiality management requirements. The sources do not say where data is stored, whether data leaves a customer site, or what cross-border data arrangements are available.

Sources

• 36kr 获达晨财智、元禾璞华数千万投资，这家物理AI企业自研WMM世界机理模型丨36氪首发 https://36kr.com/p/3981147673345033
• 36kr 产品已批量进入全球头部Tier 1产线，「知来具身」要为工业场景打造"开箱即用"的泛化机器人 https://36kr.com/p/3985410714581760
• geekpark 特斯拉、Figure 还在攻克量产，小鹏机器人已经走下产线 https://geekpark.net/news/370421
• 36kr 硬氪首发｜东莞人工智能研究院孵化，瞄准模具制造领域物理AI研发，这家企业获创新工场数千万种子轮融资 https://36kr.com/p/3963006194728324
• woshipm 物理 AI 的大结果何时到来？ https://woshipm.com/embodied/6465488.html
• ifanr 对话蚂蚁灵波 CEO 朱兴：机器人还吃不了「粗粮」 https://ifanr.com/1680322