“Big data is the driving force behind machine learning.” That conclusion from artificial intelligence expert Li Fei-Fei has gained traction among some AI practitioners. One of them is Huang Qingqiu, a former member of Huawei’s “Genius Youth” program and now CTO of Morphi Robot.
After Morphi Robot raised more than RMB 1 billion (USD 148 million) in angel funding, 36Kr met Huang in Shanghai. “If data quality were infinitely high, even a very simple model could be trained to produce excellent results,” he told the media outlet.
Amid the ambitious narratives surrounding embodied intelligence startups, Huang’s data-centric approach sounds unusually direct and pragmatic.
Huang was part of the second cohort of Huawei’s “Genius Youth” program. He joined Huawei’s automotive business unit in 2020 and became head of AI for autonomous driving. During his time at Huawei, he led efforts to apply AI to LiDAR (light detection and ranging), perception, then to sensor fusion perception systems, and eventually, during the ADS 4.0 era, to Huawei’s entire assisted driving system.
He also led the team that built Huawei’s autonomous driving data engineering system from scratch. The effort began during the ADS 2.0 era and was already fairly mature by ADS 3.0, he said.
During ADS 2.0, the data system Huang led was used only to support Huawei’s general obstacle detection network. By ADS 3.0, it was supporting all static perception tasks, traffic lights, and prediction. By ADS 4.0, the system could support the entire end-to-end stack.
That deep involvement in data engineering gave Huang a methodology he now considers invaluable as he builds a company in embodied intelligence.
Nearly every entrepreneur in embodied intelligence agrees that the endgame is large-scale adoption in homes, according to 36Kr. One major reason today’s embodied robots have yet to reach that point is that embodied models still lack sufficient generalization. Put differently, the robot’s “brain” is not yet intelligent enough.
To solve that problem, Huang sees data engineering as the key breakthrough.
“The simpler the model architecture, the better,” he said. “If data quality were infinitely high, even a very simple model could be trained to produce excellent results.”
Huang explained to 36Kr that, at its core, a model compresses information and converts between modalities. Its capability ceiling depends on its parameter count. Most improvements to model architecture, he argued, are primarily about making the model more efficient at absorbing data and have relatively little effect on its ultimate capability ceiling.
That efficiency, he added, can be offset with more data and longer training time.
That is why Huang believes that, at this stage, when high-quality embodied intelligence data remains scarce, data engineering should take priority over innovation in model architecture.
Based on that view, Huang has built a data flywheel for Morphi Robot.
For data collection, Morphi Robot developed its own lightweight wearable devices and hired cleaners in hotels, mixed-use residential and commercial apartments, and other commercial environments to collect data.
“Once the data comes back, you need to put it through rigorous quality filtering,” Huang said. “Some data may contain inaccurate movements, improperly worn equipment, or all kinds of other problems. You have to identify the valid data.”
“After that, you still need to classify it. Even if I have a massive amount of data, I need to be able to retrieve precisely which data should be used to train each model.”
Once categorized, the data is automatically labeled with the ground truth required for training, then fed into the model. Morphi Robot has also designed multiple layers of evaluation to test models at scale.
“After evaluation, for scenarios where the results aren’t good enough, we go back and collect and mine the data again,” Huang said.
Huang acknowledged that his philosophy of data engineering is hardly unique. He expects more embodied intelligence practitioners to recognize that the field “needs a closed-loop data system similar to the one used in autonomous driving.”
The industry’s data collection methods are also shifting from teleoperation toward portable collection devices, he said, because teleoperation is useful for quickly creating demos in specific scenarios but offers limited generalization.
That does not mean every company can build a good closed-loop data system.
Huang told 36Kr that the quality of such systems can come down to the details.
As an example, he pointed to “relocalization” in smart driving.
Relocalization refers to situations where two vehicles pass through the same location. After ground-truth annotation has been completed for the first vehicle, a company would ideally like the second vehicle to reuse the previously annotated data. But that is not always possible.
“You need the positioning accuracy of both the second vehicle and the first vehicle to be high enough to ensure that the two can be matched precisely,” Huang said. Many companies, he added, “may not be able to achieve centimeter-level relocalization.”
A closed-loop data system contains “tens of thousands of details like this,” Huang said.
“Every detail has to be pushed to the limit, and every detail has the potential to create a gap between companies.”
Huang’s understanding of data engineering was shaped by his experience in autonomous driving. In his technical roadmap, the closed-loop data system serves as the foundation, while developing a native embodied model in-house is the next step for maximizing the value of that data.
When it comes to model architecture, Huang does not want to label Morphi Robot’s approach as either a vision-language-action (VLA) model or a world model.
He believes two questions matter more: what goes into the model, and what comes out? And does it follow an end-to-end training paradigm?
As long as those fundamentals are in place, he argues, training methods can be changed at any time. Whether it is a VLA model or a world model, “they are essentially just tools and plug-ins for a particular stage.”
Huang’s view is that Morphi Robot will ultimately have to conduct its own pretraining and build a native embodied model.
Before reaching that stage, however, Morphi Robot first used post-training to quickly accumulate experience with physical robots and build system capabilities.
“That stage is already behind us,” Huang said.
Morphi Robot is now conducting pretraining based on open-source models. Around the end of this year, once it has accumulated enough data, the company plans to begin training its own native model from scratch.
In autonomous driving, an industry widely regarded as closely related to embodied intelligence, leading companies such as Momenta have provided some evidence that data engineering can serve as a path toward greater intelligence.
To determine whether the same setup can be replicated in embodied intelligence, more startups will have to put it to the test.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Fan Shuqi for 36Kr.
Note: RMB figures are converted to USD at rates of RMB 6.76 = USD 1 based on estimates as of August 11, 2026, unless otherwise stated. USD conversions are presented for ease of reference and may not fully match prevailing exchange rates.