PetroChina and Unitree Partner in Robotics JV, Kunlun Takes 60% Stake
A state energy giant's move to control a humanoid robotics venture with Unitree Robotics and Fanjie Intelligence signals a strategic land-grab: whoever sets oilfield acceptance standards controls the procurement pipeline.
Beijing's market regulator has disclosed a joint-venture filing showing China National Petroleum Corporation's manufacturing arm, CNPC Kunlun Manufacturing, alongside its coordinated actor Kunlun Capital, is partnering with Unitree Robotics and Beijing Fanjie Intelligent Technology to establish an industrial embodied-intelligence robotics solutions company. The equity split — 60%, 20%, 20% in favor of the Kunlun bloc — leaves no ambiguity about where operational authority sits.
The timeline underscores the urgency. Kunlun Capital completed a RMB 136 million (US$18.9 million) strategic placement in Unitree's STAR Market IPO just three months before this filing. The pivot from passive financial investor to controlling joint-venture partner within a single quarter suggests CNPC is accelerating a broader consolidation of its fragmented robotics capabilities into a single accountable legal entity — one that can write, and enforce, its own procurement specifications.
Fanjie Intelligence's 'World Model' Targets the Gap That Pure Machine Learning Cannot Fill
Unitree needs little introduction in industrial robotics circles. Its B2-Ex explosion-proof quadruped robot — rated for natural gas gathering stations and hydrocarbon processing zones — has an auditable commercial track record. The less familiar variable is Fanjie Intelligent Technology.
Incorporated in April 2026 with registered capital of RMB 1.1111 million (US$154,000), Fanjie was incubated by the team of Academician Qiao Hong of the Chinese Academy of Sciences' Institute of Automation; Qiao holds a direct 9% equity stake. The company's technical thesis centers on "world model and embodied foundation model integration" — a physics-causality-aware AI architecture that constrains algorithmic outputs via mechanistic models rather than pure statistical training. CEO Han Jinchi comes directly from Qiao's research group.
The industrial logic is straightforward. Oil and gas field operators cannot catalogue every failure mode: catastrophic leaks, flash fires, and compound emergency scenarios are, by definition, rare and long-tail events. Pure learning-based models face fundamental generalization constraints in low-sample, high-consequence environments. Fanjie's approach — if validated in field conditions — addresses precisely this brittleness. The joint venture's equity structure effectively co-locates algorithmic explainability with on-site operational authority inside a single corporate vehicle, compressing the cycle from prototype to certified acceptance.
The critical caveat: Fanjie's field-validation hours in live hydrocarbon environments are, at present, effectively zero.
CNPC's Existing Robot Fleet Reveals a Standardization Problem, Not a Technology Gap
A material fact often omitted from coverage of this deal: CNPC's internal robotics deployment is already substantial. The gap is not capability — it is interoperability.
In Xinjiang, Kunlun Digital Intelligence's explosion-proof wheeled inspection robot "Xiaozhi" holds Ex d IIC T6 certification and logged 216 consecutive operating hours in temperatures reaching minus 40°C during the 2026 Lunar New Year holiday period, substituting for more than 350 manual inspection shifts with a reported recognition accuracy exceeding 99.5%. Separately, the CNPC Pipeline Bureau's Communications Company deployed an explosion-proof quadruped robot at the Xiguo Oil Depot in May 2026 — the first quadruped inspection deployment in CNPC's oil sales segment — resolving signal-obstruction, trench-crossing, and intrinsically safe circuit-retrofit challenges in the process.
These are production deployments, not proof-of-concept demonstrations.
The structural problem is fragmentation. Xiaozhi's wheeled chassis underperforms in terrain marked by dense pipeline trenches and uneven desert topography. The Pipeline Bureau's quadruped pilot is a project-based arrangement, with hardware, algorithms, and sensing equipment sourced from multiple vendors — driving high system-integration and data-unification costs. Kunlun Digital Intelligence's core competency is software, large models, and systems integration, not equipment manufacturing. Kunlun Manufacturing, established in Xiong'an in October 2023, handles intelligent upgrades of energy equipment but lacks the AI-model layer.
The 60%-controlled joint venture is architecturally designed to occupy precisely the white space between these two units: a single product definition incorporating hardware body, AI model, and oilfield acceptance criteria — a standard solution that can be replicated at scale and written into procurement catalogues, rather than custom-assembled at each facility.
Internal Competition and Multi-Year Qualification Cycles Pose the Steepest Hurdles
The joint venture's most formidable near-term competitors are not external. CNPC's ecosystem already contains Kunlun Digital Intelligence's wheeled inspection robots, the Pipeline Bureau's quadruped trials, and accumulated third-party installations — including Qiteng Robotics' explosion-proof inspection systems, which have surpassed 8 million cumulative operating hours in industrial environments. All of these compete for the same capital expenditure budget.
Oil and gas procurement follows maintenance-regulation-driven logic. A robot that completes pilot acceptance is not, thereby, a standard procurement item. The qualification pathway — from field trial to framework purchase agreement — typically spans multiple years of stability assessment. Crucially, a technology does not enter routine procurement until it is written into formal operating and maintenance procedures. The 60% controlling stake confers product-definition authority; it does not accelerate the accumulation of field operating hours.
The international benchmark is instructive. ANYbotics' ANYmal X required four years of engineering development and more than ten patents before delivering a product simultaneously certified to ATEX and IECEx standards — with Equinor as first recipient. BP, Shell, Petronas, and Petrobras have been conducting early-access testing across hundreds of onshore and offshore facilities since 2023.
One specific physical constraint surfaces immediately in the CNPC context. Unitree's B2-Ex carries a nominal operating temperature floor of minus 20°C. Xinjiang oilfields routinely record minus 40°C in winter — the same environment where Xiaozhi operates on purpose-built low-temperature batteries with full cold-chain-hardened system design. Winter operability may represent the joint venture's first mandatory product-specification gap to close.
The Verdict: Certification Is the Entry Ticket; Operating Data Is the Only Currency That Matters
The strategic logic of Kunlun Manufacturing's 60% controlling stake is coherent: consolidate dispersed robotics capabilities, internalize acceptance-standard authority, and prevent external vendors from defining what constitutes a qualified oilfield robot. The structure positions CNPC to shape — not merely consume — the industrial embodied-AI product category.
Execution risk, however, is concentrated in a single variable. The joint venture must generate sufficient cumulative field hours — combining Unitree's hardware reliability data with Fanjie's model interpretability logs — before Kunlun Manufacturing's acceptance standards become bankable procurement criteria rather than corporate aspirations.
The practical test is specific: after the first batch of Fanjie-model-equipped Unitree B2-Ex units completes a full winter operating cycle at a CNPC oilfield, does the resulting run report get incorporated into the next round's technical specification documents? Standards cannot generate data. Data, however, can define standards — and right now, that data does not yet exist.
Related Coverage:
Unitree Robotics, Sequoia China Launch Joint Fund as Embodied AI Makers Pivot to Venture Capital
Unitree's Wang Xingxing Sees Physical AI Breakthrough Within Two Years