Robots can dance, perform backflips, and run marathons, yet many would struggle to unscrew a bottle cap they've never encountered. To bridge this final gap into the physical world, what's needed isn't just a more agile body, but a brain that's been saturated and repeatedly tempered by high-quality data.

From freezing ice-covered roads to scorching desert sands, from mirror-smooth glass surfaces to rugged mining trails, robots can undergo tens of thousands of parallel training sessions and learning cycles under extreme conditions, achieving self-evolution. However, none of this unfolds in the real world—it all happens within a 'digital parallel universe.' The person building this digital 'academy' for robots is Yang Haibo, co-founder and president of Guanglun Intelligent, along with his team.

'We want robots to learn knowledge in classrooms, test their abilities in exams, and identify weaknesses during real-world deployment—just like humans—forming a continuous learning loop,' says Yang Haibo.

Pioneering the Data Frontier of Physical AI

At the 2026 NVIDIA GPU Technology Conference, the tech giant's CEO Jensen Huang showcased simulation scenes of robots folding clothes and tightening belts during his keynote. The technology traces back to a Beijing-based Chinese company: Guanglun Intelligent. The story began three years ago when large language models were the hottest trend, with parameter scale and algorithmic innovation dominating the spotlight. Beijing native Yang Haibo, however, spotted a different direction—the 'data scarcity' of the physical AI era.

'Large language models are like self-taught prodigies, learning from vast internet text accumulated over decades. Robots can't do that; what they need to learn is grasping, pushing, collision, and deformation—physical interactions with no ready-made textbooks,' explains Yang. Robots don't lack smarter algorithms; they lack a complete learning system: someone to teach them, test them, and help correct their mistakes—just like human growth, where classes, exams, internships, and reviews are all essential.

Based on this, robots need more than a set of physics textbooks—they need a 'school' where they can repeatedly attend classes, practice, take exams, and review their performance. 'Just as pilots train in flight simulators, robots can make all their mistakes in a virtual environment before working in the real world—then they're solid,' says Yang.

In early 2023, Guanglun Intelligent was established in Haidian District, with its first use case targeting four-wheeled robots: autonomous vehicles. However, simulation data was still a niche market at the time, and some investors expressed deep skepticism upon hearing the project involved 'simulation.'

'The autonomous driving industry placed greater emphasis on real-world road test data. In some people's eyes, simulation data was merely a nice-to-have—its true value wasn't yet recognized.' But this underestimated potential was quickly validated in a tough battle. In the second half of 2023, Guanglun Intelligent secured an order from an automotive OEM, with a contract clearly stipulating: the company's data must improve the client's model accuracy to a specific percentage point. This requirement was extremely demanding because once data is delivered, how, when, and with what parameters the client trains their autonomous driving model is entirely outside the supplier's control.

'We gritted our teeth and pushed forward!' Through relentless day-and-night efforts, a city 'generated' by algorithms emerged from nothing: streets and traffic lights were gradually built, along with stormy nights, suddenly braking vehicles ahead, and pedestrians crossing roads—every second generating new driving scenarios. The autonomous driving system practiced and evolved repeatedly in this environment, ultimately earning client recognition and laying a solid foundation for the team's journey into the embodied intelligence track.

Self-Developed Breakthroughs Build a Robot 'Academy'

In a new energy vehicle factory, a robot is learning a precision task: assembling battery modules. The robot's vision sensors quickly scan and identify part specifications; the robotic arm adjusts its angle and approaches along an optimal trajectory; the flexible gripper gently grasps, with real-time force feedback fine-tuning to ensure the part is neither crushed nor dropped, before precisely and smoothly inserting it into the slot—the entire sequence flows seamlessly.

It's hard to imagine that this 'skilled worker' completed its training and went live in just one week. Previously, such training could take months, requiring dedicated production lines, vast quantities of real parts, and constant engineer supervision. Every collision meant stopping production to replace parts, with robots needing to trial-and-error hundreds of thousands of times. Now, within Guanglun Intelligent's 'digital academy,' parts can be replicated infinitely, and even real-world production anomalies like deformation and deviation can be simulated.

According to estimates, within this 'digital academy,' companies can train robots with dramatically improved efficiency, at costs reduced to one-tenth or even lower. In 2024, the embodied intelligence wave swept in. Rather than directly building robots or large models, Guanglun Intelligent chose to stand at the edge of the spotlight as the 'pick-and-shovel seller' behind the tech gold rush.

'By conservative estimates, the high-quality data gap facing embodied intelligence is at least 1,000 times larger than that of autonomous driving,' says Yang. Facing this vast blue ocean, the team chose a harder path: full self-development. At the time, some simulation tools existed in the market, but game engines prioritized visual flair over physical realism, traditional industrial simulation software was too slow to support the massive parallel processing required for AI training, and closed underlying architectures couldn't adapt to diverse domestic robot hardware and computing ecosystems.

'Only by self-developing could we deeply optimize the entire data production pipeline based on the real needs of robot training.' This near-obsessive persistence allowed Guanglun Intelligent to avoid being held back by previous-generation technological paradigms and instead redefine how data is produced. Take the industry-level challenge of 'flexible cable assembly': cables are soft and prone to deformation, with extremely complex force dynamics. Through its self-developed solver and high-fidelity measured parameters, the team successfully simulated precise deformation and multi-dimensional force feedback of cables under various grasping, twisting, and plugging states in a virtual environment—enabling robots to reliably acquire these skills and ultimately transfer them efficiently to real production lines.

The architecture of the robot 'academy' has gradually taken shape: EgoSuite, a data solution providing high-quality human-experience 'textbooks' for robots and world models; RoboFinals, an industrial-grade simulation evaluation platform serving as the 'exam hall' for embodied intelligence; and RoboStack, an industrial-scale evaluation platform that helps robots 'get employed' while collecting new experiences and failure cases for continuous learning and evolution.

Superior technology translates into market recognition. After the 2025 Spring Festival, Guanglun Intelligent successively secured orders from three of the most prominent model developers and hardware manufacturers in embodied intelligence. More important than the amounts: this time, they no longer needed to self-validate data quality—they directly delivered data based on their own standards and received tangible 'votes of confidence.'

As the embodied intelligence industry hits the inflection point from 'showcase' to 'application,' market demand is being released exponentially. Since the start of this year, orders for the company's data and evaluation services have far exceeded expectations—revenue grew 10-fold last year, and first-quarter revenue this year has already surpassed the full-year total of last year. Today, many of the world's leading embodied intelligence teams source their simulation assets and synthetic data from Guanglun Intelligent. Meanwhile, the company has achieved scaled delivery across three dimensions—simulation synthetic data, simulation evaluation, and first-person human video data—with delivery volumes leading the industry.

Chinese Companies Move to the Center of the Table

Alongside commercial breakthroughs, Guanglun Intelligent has begun climbing to higher and deeper levels of the industry. In July of this year, the company was invited to join the EgoVerse International Embodied Human Data Committee, working alongside Meta, Scale AI, and other international teams to build the world's largest human data training system. It is currently the only Chinese company on the committee. Just months earlier, Guanglun Intelligent had been invited to join the technical steering committee of Newton, the international open-source physics simulation engine, to drive technological governance, ecosystem collaboration, and international standard evolution for next-generation physical AI simulation infrastructure. Seated alongside them were four of the world's top institutions: NVIDIA, Google DeepMind, Disney Research, and Toyota Research Institute.

Previously, global technical leadership in physics simulation engines rested entirely in the hands of American and European companies. Now, both Newton and EgoVerse—two international standards leading next-generation open-source physical AI—feature deep participation from Chinese companies. 'We're not building a data company; we're building a sustainable 'school' for the entire industry. Whoever needs to train robots can come here,' says Yang.

While stepping onto the international stage, Yang also sees a deeper concern: in the physical AI era, simulators and solvers are infrastructure as critical as computing chips, yet they represent the weakest link in China's physical AI landscape. High-end industrial simulation has long been dominated by overseas companies, which have built lasting advantages in underlying technology, engineering processes, standards systems, and ecosystem rules. Embodied intelligence's reliance on simulation runs deeper and broader than industrial simulation—it isn't just a design tool; it determines the ceiling of training and evaluation capabilities for the robotics industry. If underlying engines, key standards, and evaluation systems solidify abroad, China's embodied intelligence industry will face a long-term constraint of 'models can run, but training and evaluation environments are controlled by others.'

Driven by this thinking, the team is fully committed to self-controllable simulator development, building a full-stack simulation platform integrating 'solver-measurement-generation.' Its self-developed high-precision GPU physics solver features differentiable, multi-physics, multi-material unified solving capabilities, supporting high-precision real-time simulation of complex physical processes involving rigid bodies, soft bodies, fluids, and granular materials. While sharpening internal capabilities, a collaboration network is also being woven: partnering with Moore Threads to build 'domestic computing + self-developed simulation' self-controllable embodied intelligence infrastructure; collaborating with international giant Siemens to break down barriers between industrial simulation and embodied simulation; and jointly open-sourcing the Isaac Lab-Arena embodied evaluation benchmark framework with NVIDIA.

Every thread of this network answers the same question: when the physical AI era arrives, can China secure a voice in underlying infrastructure and remain at the center of the table? Yang is confident. 'Beijing's talent advantages are unique—not only does it gather the highest concentration of top AI talent, but within a ten-kilometer radius of the company, we can find nearly all of our collaborative partners,' Yang says with admiration. 'This is the best place to realize dreams.'