Embodied AI Goes Global to Close the Data Loop: Imperial College Postdoc's UK Startup Raises Tens of Millions of RMB

"A grounded sample of embodied globalization." By Qian Ren

While numerous domestic embodied intelligence companies grapple with data anxiety and rush to build data collection factories, a British company founded by a Chinese entrepreneur is already validating a fundamentally different scene-data closed-loop path in the world's factories — with Europe as its proving ground.

"Dark Waves" has learned exclusively that Extend Robotics, a vertical-scene embodied service provider, recently completed a Pre-A round of tens of millions of RMB across China and the UK. Investors include Dreame Ventures, Zhuopu Investment, and European firm Neo Venture.

CEO Chang Liu is a postdoctoral researcher in drone technology at Imperial College London. His research focuses on the convergence of robotic teleoperation, human-robot interaction, and virtual reality. During his doctoral studies, he specialized in autonomous flight and environmental perception algorithms for aerial robots, developing VR-based intuitive robot control systems that enable non-professional users to remotely operate robotic arms. He has published five papers in IEEE journals.

In an era of breakneck progress in embodied intelligence, Extend Robotics has carved out a niche in the interstitial market between China and overseas — finding an entry point for embodied systems to acquire real-world scene data. European factories face severe labor shortages, yet local manufacturing still contains vast pockets of small-batch, high-flexibility operations that resist automation upgrades, making it the most suitable market for embodied intelligence to land in real scenarios. Chinese embodied hardware manufacturers may hold product advantages, but they struggle against invisible barriers in the European and American markets: scene services, local trust, and data compliance.

Extend Robotics plays the pivotal role of breaking through this bottleneck. It doesn't build hardware itself. Instead, through its proprietary 3D remote transmission compression algorithm and zero-barrier VR teleoperation system, plus a semi-automated deployment and leasing business model, it embeds Chinese hardware into European industrial scenes. It primarily provides embodied robot services for high-risk scenarios and complex industrial environments in Europe, serving over 30 paying clients including Leyland, Airbus, and Scale.ai, and has been officially certified by NVIDIA as one of Europe's top five embodied intelligence companies. Recently, Extend became an official service provider for Unitree Technology, partnering to accelerate the deployment of humanoid robots in real-world scenarios.

Its technical edge lies in the deep integration of consumer-grade VR, AI, and robot control. Through a 3D immersive teleoperation interface, personnel without technical backgrounds can intuitively operate robotic arms and humanoid robots to perform precision tasks. Meanwhile, the smooth, high-quality real-machine data collected during operations trains its own models for generalization, improving accuracy and cycle time — forming a closed loop of "remote operation → data collection → intelligent iteration."

The significance of Extend Robotics as a case study may be this: It proves that there is no "endgame" theory for the landing of embodied intelligence. Finding solutions to structural pain points in the real world is the necessary innovation on the path toward any endgame.

Part 01

Pivot and Survival: From Drones to Service Provider

When Chang Liu decided to start his company in 2021, his initial vision was to equip drones with robotic arms for aerial teleoperation. But Europe's stringent aviation regulations (CAA) quickly doused that plan. Facing regulatory barriers, he judged that while a demo was feasible, commercialization would be an uphill battle. He made a pivotal decision: cut hardware manufacturing, go all-in on software. Leveraging the UK's unique ecosystem — strong in software, high in hardware costs — Extend began refining its VR-based remote operating system, converting its early experience with 3D-printed robotic arms into standardized capabilities for interfacing with third-party hardware.

This pivot established Extend's underlying business logic: don't build heavy-asset hardware bodies, build light-asset "scene integration" and "human-robot interaction middleware." Through its self-developed 3D Mesh compression algorithm, Extend achieved a technical breakthrough — compressing 3D scene data to 1/100 of its original size with latency controlled under 50ms. This directly solved VR teleoperation's biggest pain point: motion sickness.

In traditional solutions, operators would often vomit after half an hour of continuous operation. Extend's solution enables prolonged use. This technical approach allowed Extend to interface with various third-party robotic arms (including hardware from Chinese manufacturers), making it one of Europe's earliest embodied intelligence companies to achieve commercial deployment.

From drones to screwdriving, Chang Liu sees no demotion — just the same problem being solved: how to extend human capabilities beyond physical limits.

After the pivot, Chang Liu quickly realized that what truly made clients pay wasn't how advanced the technology was, but whether it solved real problems. Post-Brexit, UK manufacturing faced severe labor shortages. During client visits, he found that production lines were stuck on fine-manipulation steps that resisted automation. Agricultural pickers were also draining away; vineyards had only 2-3 week harvest windows, and losses from rotting grapes far exceeded labor costs. At auto plants, high-voltage battery installation positions paid annual salaries of 400,000-500,000 RMB and still couldn't find takers.

Traditional industrial robot giants like KUKA and ABB excel at fixed-programming, large-scale repetitive operations. But faced with small-batch, multi-variety, high-flexibility rework processes, as well as high-risk scenarios like nuclear handling and space maintenance, they're helpless. These scenes don't need "automation" — they need "human-robot collaboration," letting humans remotely control robots to complete work that machines can't handle alone.

This is precisely Extend's opportunity.

Unlike domestic embodied intelligence companies generally pursuing L4 full autonomy and investing heavily in building data collection centers, Extend chose a path seen as a "transitional solution": using L2/L3 semi-automation (human teleoperation + AI assistance) as the entry point, generating cash flow through subscription models, collecting effective data from real scenes, then gradually improving model capabilities to evolve toward full autonomy.

This strategy stemmed from an accurate reading of European industrial reality. Extend fills exactly this gap market. In the Leyland Trucks case, Extend deployed a semi-humanoid robot solution for high-voltage battery module installation, replacing manual operations that previously required heavy protective suits. This wasn't simple hardware-software sales, but a "turnkey" service encompassing hardware selection, production line debugging, compliance certification, and ongoing maintenance — priced at 1-2 million RMB per project, with clients able to break even within 18 months.

More strategically valuable is the positive accumulation of data assets. Through the teleoperation process, Extend collects large volumes of edge-case data from real scenarios — robotic arms jamming, sudden lighting changes, human intervention. This isn't lab-generated data; it's real-scene data with genuine physical constraints. Chang Liu notes that domestic companies spend enormous sums building data collection centers and paying daily for data acquisition, while Extend's service-collected data belongs to the company, with users paying to collect it. When operators remotely handle abnormal working conditions, this data automatically flows back to train SOTA models, gradually improving automation levels and forming a "data flywheel." This model lets Extend accumulate data while making money, creating a virtuous cycle.

Currently, Extend has accumulated 35 subscription clients, spanning automotive manufacturing (Paccar, Ford, JLR), aerospace (Airbus, UK Space Agency), nuclear energy (AtkinsRéalis, INL), robotics (Unitree Technology), and agriculture (a grape-picking project in collaboration with Queen Mary University), with revenue growing 100% year-over-year for the past three years.

Part 02

Reconstructing the Globalization Path for Chinese Embodied Intelligence

Chinese hardware manufacturers like DOBOT, RealMan, Unitree Technology, Star1, and AgileX Robotics possess the world's most cost-competitive robotic arms. But when facing the European market, they get stuck on CE certification, safety compliance, GDPR data sovereignty, and building local client trust. These hurdles are nearly impossible to clear alone. They don't lack technology or products; they lack the entity that can get their products into the European market.

As a service provider, Extend shoulders these high-cost, high-barrier landing operations. The company employs a "dual-track" organizational structure: the China team handles core algorithm iteration and rapid response, while the UK team manages compliance, client relationships, and local operations. This architecture lets Extend simultaneously satisfy Chinese supply chain efficiency demands and European factory landing requirements.

According to Chang Liu, when Extend helps Chinese hardware manufacturers enter the European market, it even prepays 50% of CE certification costs. These aren't trivial sums — a full certification process runs from several hundred thousand to over a million RMB. But Extend is willing to bear this cost in exchange for subsequent long-term orders.

The "software-hardware integration + deep integration" model is reconstructing the globalization path for Chinese embodied enterprises. Extend has formed strategic partnerships with DOBOT, RealMan, and others, helping their hardware pass admission validation for Europe's extreme scenarios, while providing Chinese manufacturers with local scene data and iteration feedback. In this relationship, Extend is both a software provider and a "critical bridge" for hardware going overseas.

However, ecological niche risks remain. As an "official partner" in NVIDIA's Isaac ecosystem, Extend both benefits from NVIDIA's technical support and faces the risk of being marginalized by a platform giant — when NVIDIA's GR00T blueprint matures and teleoperation data standards are unified, Extend could be downgraded from "solution provider" to "data supplier." This dependence on platform giants is a risk all ecosystem companies must confront.

Meanwhile, as Extend establishes deeper cooperative relationships with domestic hardware partners, new questions arise: how to maintain "hardware neutrality"?

Chang Liu's answer: dig deep into fields with sufficiently high barriers — nuclear handling, space maintenance — building irreplaceable industry expertise in these areas. These domains have long certification cycles, high safety requirements, and strong client stickiness; once you're in, you're hard to displace. Simultaneously push software platformization, shifting from selling integrated hardware-software solutions to building training and data platforms. This way, even if future competition emerges from non-partner hardware, Extend can maintain bargaining power through data assets and platform capabilities.

In interviews, Chang Liu offered grounded feedback on the current "general large model" frenzy in the embodied intelligence industry, from a practical deployment perspective. Cutting-edge general embodied models may look stunning on paper, but get to Leyland Trucks' shop floor and a screw's tolerance deviation, a stray reflection, can bring the most impressive model to its knees. Real-scene data is an indispensable link on the road to pursuing AGI's "grand narrative." Acknowledge that general AI can't yet handle the physical world, and first achieve real data closed loops within scenes through teleoperation technology.

This pragmatism stems from understanding the laws of technical evolution. Referencing autonomous driving's decade-long journey still without achieving full driverlessness, Extend judges that embodied intelligence will enter a prolonged "human-robot collaboration" phase. L2-level human teleoperation intervention will persist for a long time, gradually transitioning to L3/L4 limited-scene autonomous operation through continuous accumulation of real-scene edge-case data.

Beyond its integrated hardware-software + teleoperation data collection solution, Extend is also exploring a leasing model with higher client acceptance and broader industry coverage. Users need not purchase the full machine, but instead receive automation transformation in a manner akin to hiring labor. A Result-as-a-Service pricing model greatly reduces client concerns about cycle time and success rates, while allowing Extend more flexible access to stable cash flow and larger-scale industry data at no cost — laying groundwork for autonomous embodied model training.

To date, Extend Robotics has validated the feasibility of its business model: 35 paying clients, coverage across four sales regions (Europe, America, China, UK), and a triple revenue structure of software + hardware + services.

In the long cycle of embodied intelligence, this company has proven one thing: technological breakthroughs' disruption of industry requires "practitioners" leaving footprints step by step. The companies that solve concrete problems in concrete scenes, build geopolitical bridges, and get Chinese hardware actually running in European workshops — these may be the final survivors.

Layout by Zhixin Han

Images from Visual China

Recommended Reading

The flow of money, the rise and fall of people