From Huawei Autonomous Driving to General-Purpose Robotics: Huang Qingqiu Wants to Rebuild a System for the Physical World | BlueRun Ventures Family Headlines
From Single-Demo to General-Purpose Robot: The Critical Leap
"In the lab, model architecture sets the ceiling. But in the real physical world at million-unit scale, systems decide survival."
This is how Huang Qingqiu, co-founder and CTO of Moqi Intelligence, views embodied AI.
It is precisely this conviction that has helped Moqi Intelligence close over 1 billion RMB in angel-round financing within just six months of founding, at a valuation exceeding 7 billion RMB. BlueRun Ventures participated from the very first angel round and continued to double down in the angel+ round, making it one of Moqi's earliest and most steadfast investors.
This early conviction stems both from an understanding of where embodied AI is ultimately headed, and from recognition of the founding team's complementary strengths. Huang Qingqiu previously led Huawei's intelligent driving algorithm breakthroughs from ADS 1.0 to 4.0 and their mass production at million-unit scale. Gao Wenli brings years of overseas market expansion and global operations experience. One excels at making technology run stably in complex physical environments; the other at replicating complex systems across broader markets. This combination meant Moqi Intelligence never chose the single-point demo path from day one. Instead, it confronts the harder, longer-term challenge: using integrated hardware-software systems engineering to build toward general-purpose home robots.
We believe the future competition in embodied AI will not be about any single model, paper, or demo — but about who can build an engineering system capable of continuous self-evolution. Huang Qingqiu's framework of "real-time closed loop, hardware-software integration, data closed loop, and systems engineering" captures the central tension as embodied AI moves toward real-world deployment. Models are merely the starting point; real-world data loops, hardware-software coordination, engineering reliability, and scalable mass production are the true watersheds that will determine industry structure. We look forward to more good news from Moqi Intelligence as it accelerates embodied AI's transition from model innovation to systems innovation, from laboratory to industrial reality.
Below is an in-depth interview with Huang Qingqiu by An Yong Waves — enjoy.
"In the Physical World, Systems Decide Survival."
By Ren Qian
Edited by Ba Rui
An Yong Waves has learned exclusively that embodied AI company Moqi Intelligence, founded just six months ago, has completed over 1 billion RMB in angel-round series financing within six months. The company is valued at more than 7 billion RMB, representing one of the larger disclosed first-round funding rounds in China's embodied AI track to date.
According to An Yong Waves, investors in this round include Alibaba, Tencent, BlueRun Ventures, Legend Capital, Hua Capital, Gaorong Ventures, CASSTAR, Source Code Rhythm, Luminous Ventures, Bairui Capital, and 58 Industry Fund.
Moqi Intelligence was co-founded by Huang Qingqiu and Gao Wenli.
CTO Huang Qingqiu, born in 1994, is the archetypal "technical founder": undergraduate degree in automation from Tsinghua University, PhD from The Chinese University of Hong Kong's MMLab under Dahua Lin, with long-term research in robot control, computer vision, and AI algorithms, and dozens of top-conference papers to his name.
In 2020 he joined Huawei's Intelligent Automotive Solution Business Unit through its "Genius Youth" program, later becoming head of the intelligent driving AI department. He led Huawei's intelligent driving algorithm breakthroughs from ADS 1.0 to 4.0 and their mass production — the first in the industry to achieve million-unit scale production of one-stage end-to-end architecture (WEWA).
CEO Gao Wenli was previously co-founder of cross-border logistics company iMile, the "key player" who built out the Middle East network and replicated operations systems across global markets. Earlier, he spent 11 years at Huawei's carrier business group, responsible for overseas market expansion.
The founders' backgrounds determined that Moqi Intelligence would not pursue single-point demos from day one, but instead anchor itself in "integrated hardware-software systems engineering capability," using commercial service scenarios as technology proving grounds, aiming for the super-terminal of general-purpose home robots.
Embodied AI may be one of the most complex systems in engineering history. Despite being an industry with hundreds of billions in capital deployed, embodied AI's iteration speed lags far behind autonomous driving and large language models — because the challenge of enabling robots to have a "brain" through "real workers performing real tasks in real scenarios" to collect data has not yet been fully cracked. But this is precisely the current core battleground: a comprehensive competition between general-purpose embodied brains and systems engineering capability.
Huang Qingqiu believes that without building the brain and doing pre-training, this wave of embodied entrepreneurship has no meaning. "With so much capital and talent attracted, if we only stay at the body or single-task level, embodied AI can hardly bear the industry's current expectations." In his view, the endgame of embodied AI lies not in whose large model has more "intelligence," but in who can build a scalable, reusable, closed-loop underlying architecture in a chaotic, unstructured physical world.
An Yong Waves met Huang Qingqiu in Shenzhen, where we discussed the industry inflection point for embodied AI, endgame judgments about home scenarios, and how intelligent driving experience transfers to embodied AI. The conversation follows.
The interview has been edited by An Yong Waves —
An Yong: New embodied AI companies keep emerging, and some share a similar vision to yours. What do you think truly matters?
Huang Qingqiu: If I had to answer in one word: systems engineering. In the lab, model architecture sets the ceiling. But in the real physical world at million-unit scale, systems decide survival.
Today the industry debates mostly VLA versus world models, end-to-end versus hierarchical, autoregressive versus diffusion. But personally I don't think these are the most critical questions.
Embodied AI is fundamentally about building a hardware-software integrated real-time closed-loop system. What we really need to do is start from first principles, fusing hardware boundary definition, rapid data flow, efficient model iteration, and closed-loop evaluation mechanisms into a self-evolving organism.
An Yong: What does a real-time closed-loop system mean?
Huang Qingqiu: Real-time and closed-loop are the essential characteristics of Physical AI. Large models "monologue" in static corpora; embodied AI "grapples" in a dynamic physical world.
For example, when a robot lifts a paper cup full of water, it must adjust force every millisecond based on cup weight, water sloshing, and finger-cup friction — lifting without crushing or spilling. That's "real-time." And every action changes the environment, which immediately feeds back to affect the next action; action and environment are mutually causal. That's "closed-loop."
"Real-time" requires operation and motion control models to run on-device, with the brain at least 10Hz and the cerebellum at least 100Hz. This frequency redline pulls models from the cloud's "greenhouse of infinite compute" back to the device's "cage of power consumption and heat dissipation," which in turn constrains compute and locks parameter count. It's not that we don't want to go bigger — physical law forbids it. "Closed-loop" means any sensor precision loss or actuator latency gets amplified by the system, which also determines what models learn and how they learn. So at this stage, hardware and software simply cannot be decoupled for iteration. The only path to rapid system iteration is hardware-software integration.
An Yong: This system is your underlying architecture logic. What about product? What is Moqi Intelligence's ultimate product positioning and endgame direction?
Huang Qingqiu: We see the endgame in the home. Moqi Intelligence aims to build general-purpose home robots. We will release our first service-oriented robot in July, taking a step forward.
An Yong: A question many people have: compared to existing robot vacuums, dishwashers, and so on, what is the necessity and core difference of general-purpose embodied robots entering the home?
Huang Qingqiu: These products are excellent, but they solve fragmented, point-like needs. And in solving problems, they often create new "human problems." Users have to rearrange furniture for them, regularly clean dust bins and swap mop pads, load dishes according to the dishwasher's logic, set electronic fences for lawn mowers. This is humans serving machines, not machines serving humans.
Robots are the optimal form to change this status quo. Moqi Intelligence's endgame product will be a home robot with global physical understanding capability that can handle complex tasks. It can recognize all changes in the home and autonomously complete cleaning, organizing, delivering, and various other tasks. It doesn't require any compromise from you — it adapts to your life order, not the other way around.
An Yong: Why are you entering embodied AI now? Some embodied companies are already actively preparing for IPO. Don't you feel it's late?
Huang Qingqiu: Not late at all. This is precisely the moment when two technological singularities converge.
The first is the GPT moment. Large language models excellently solved high-level semantic understanding and task decomposition — robots finally have a "cognitive brain." But in my view, having only a brain is insufficient. The current biggest bottleneck in embodied AI lies in low-dimensional physical execution. The second critical moment is intelligent driving's end-to-end moment. I experienced this history in the intelligent driving industry; it proved that using neural networks to drive real-time closed loops for physical entities is feasible.
High-dimensional cognitive intelligence and low-dimensional physical control — these two originally parallel lines are converging today, and their intersection is embodied AI.
As for companies preparing for IPO, that proves capital's long-term optimism about this track, but it doesn't mean technology and products have converged. Embodied AI is a systems engineering marathon; going public is just getting a supply coupon. The race has barely started.
Moreover, from day one we anchored on the general-purpose home robot endgame and the hardware-software integrated path, which is fundamentally different in underlying logic from existing players. Home robots are a trillion-level global blue ocean. We built our architecture for globalization from Day 1, so there's no "too late to enter." Getting the underlying system and product right matters more than rushing for short-term milestones.
An Yong: But many people see autonomous driving and embodied AI as two separate tracks. You come from intelligent driving. How do you view the relationship, and to what extent can intelligent driving experience transfer?
Huang Qingqiu: I see intelligent driving as a subset of embodied AI. Human operations in the physical world can be summarized as movement and contact. Intelligent driving handles movement in a 2D plane; embodied AI handles movement and contact in a 3D world. The shift from 2D to 3D means your solution space explodes. You need to contact objects, to deeply understand the physical properties of all kinds of objects. This is also embodied AI's greatest challenge relative to intelligent driving.
Intelligent driving, as a simplified version of embodied AI, offers much experience to draw upon. End-to-end technical methodology, data closed-loop iteration systems, on-device model real-time optimization, multi-sensor fusion engineering solutions, even million-unit scale quality control processes — these are proven mature experiences from intelligent driving that can be borrowed in embodied AI R&D, substantially shortening the industry's cycle from demo to mass production.
An Yong: What difficult problems do entrepreneurs in the industry generally underestimate?
Huang Qingqiu: First, data quality. Future embodied foundation models depend on scalable high-quality data. How to collect, filter, and process high-quality data is a systematic challenge.
Second, long-term stability of embodied systems. In unstructured worlds, environments are dynamic and full of perturbations. A system that "can run" doesn't mean it "can run at scale." The gap between 1 and 1 million units shipped is a massive engineering chasm.
An Yong: Everyone is now collecting and training data through various paths. Is this the industry's solution?
Huang Qingqiu: Data is certainly most important, but it depends what data you're stockpiling. I firmly believe that the endgame form of embodied data must satisfy "three reals": real workers, in real scenarios, doing real work.
The endgame should have two collection devices: one purely for video, extremely lightweight, perhaps just a pair of glasses, collecting massive video for pre-training; the other an improvement on existing portable devices, finding the sweet spot between wearability and reconstruction precision, collecting relatively smaller amounts of data for both pre-training and post-training — both must satisfy the "three reals."
Additionally, data quality far outweighs quantity. Getting sub-millisecond time synchronization, high-speed motion trajectory precision in weak-texture environments, and so on right, ensuring every hour of data is high-quality, matters far more than quickly spreading out to collect tens of millions of hours.
An Yong: What path do you follow for models?
Huang Qingqiu: Simplifying, a model is also a system, essentially doing two things: information compression and modality alignment. Embodied model inputs are high-dimensional tokens from images, touch, language — tens of millions of dimensions — while outputs are action tokens of just dozens of dimensions. This is both compression from high to low dimension, and alignment across modalities. Low output dimension, sparse feedback, and large modality differences make it very difficult to learn.
What engineering must do is find every possible way to add "multimodal dense feedback." Today's heated debates about various approaches are essentially doing this from different dimensions: having models generate video, predict future frames — dense feedback in space, i.e., world models; having models express reasoning in language — dense feedback in logic, i.e., an important training method for VLA; adding depth maps is geometric dense feedback, adding semantic segmentation is semantic dense feedback.
So world models and VLA are not in conflict at all, not mutually exclusive, but "auxiliary lines" laid from different dimensions — which brings us back to that phrase: architecture debates are a false proposition. Moqi adopts precisely a multi-expert architecture fusing various dense feedbacks, aiming to build a native multimodal model integrating "understanding — generation — decision-making."
An Yong: Developing native multimodal models from scratch and doing pre-training require sustained heavy investment. How do you think about ROI?
Huang Qingqiu: Developing a general-purpose embodied brain from scratch is indeed long-term heavy investment, but I don't think this is fundamentally a question of calculating short-term ROI.
If we only stay at the body or single-task level, embodied AI can hardly bear the expectations of today's industry investment. Long-term, true value will return to transferable brains, data closed loops, and scalable product systems.
Of course, even if we were to calculate it, I think the numbers work out. Model training investment is after all fixed cost. Facing a high-value product shipping in massive volume, any fixed cost eventually becomes negligible.
An Yong: Was choosing entrepreneurship a difficult decision?
Huang Qingqiu: Starting a great embodied AI company was, for me, a natural choice.
At Tsinghua, besides classes and sleep, I spent 80% of my time in Room 508 of the main building — Professor Zhao Mingguo's robotics lab. Those four years, I played with robots of every form there, from line-tracking cars to self-balancing cars, to joining the Fire God team building humanoid robots, and finally for my thesis working with seniors on an autonomous bicycle that could follow people around. This is where I completed my embodied AI initiation.
I also met three seniors who, following Professor Tang Xiaoou, had founded SenseTime as its first three employees. Through their referral, I interned at SenseTime. In the early days SenseTime had just over ten people, crammed into a Wenjin International Apartment in Tsinghua Science Park — the living room was the office, the bedroom was the dorm. I still miss that free, efficient, pursuit-of-excellence atmosphere. It infected everyone there at the time, so much so that colleagues who pulled all-nighters writing code together went on to found star tech companies like Momenta and MiniMax.
Having experienced that 0-to-1 fervor, the seed of "building a great tech company" in your heart can never be suppressed again.
An Yong: During your PhD you researched "using AI to analyze movies," which sounds completely unrelated to robots. What influence did it have?
Huang Qingqiu: This topic exposed me earlier to multimodal unstructured data, and trained my ability to build a system from 0 to 1.
At the time my advisor Dahua Lin told me, "Everyone is researching short videos of a few seconds. Why don't you look at what research can be done with movies, these hours-long videos?" So for four years of my PhD, my main activity was staring at thousands of movies and brainstorming topics — from actor recognition to multimodal alignment to automated editing.
This was essentially systematically decomposing problems and building small research systems. Building systems is a process of accumulation; before the system forms there may not be much output, but once the system is built, explosive growth follows.
An Yong: After your PhD you went to Huawei's vehicle BU to work on intelligent driving algorithm R&D. What did Huawei mean to you?
Huang Qingqiu: If my PhD taught me to "look up at the stars and open new frontiers," Huawei's five years taught me to "keep my feet on the ground and fill the pits," giving me the most extreme iteration thinking and quality consciousness.
Huawei has a quality management system that could be called "nightmarish." Every tiny issue is continuously tracked. Before every new version you face soul-searching questions: "Why wasn't this problem solved again?" The process is painful, but it is precisely this refusal to let any small problem go, this grit to grind through version after version, that iterates excellent products. This accumulation gives me sufficient confidence today to build excellent embodied products, to turn fanciful technology into engineering deliverables that never fail in the physical world.
An Yong: If you don't talk about your resume, how would you introduce yourself to a stranger?
Huang Qingqiu: I'm Huang Qingqiu, a pragmatic idealist.
I believe with absolute conviction that within ten years, general-purpose robots will become humanity's "super terminal" in the physical world, entering millions of homes.
But at the same time I am even more clear that this day will not suddenly arrive because of a viral paper or a stunning demo. To cross the chasm from demo to product, there are no shortcuts. You can only trudge through the mud pit step by step: obsess over sub-millisecond multi-sensor time synchronization, scrutinize every tiny error accumulation in motor actuators, clean every trajectory anomaly in training data, catch every exception branch from loose wiring or dropped packets. The road to this super terminal is destined to be long and arduous.
The Vibe Working Era Begins! Kimi Work Beta Opens for Testing, Letting AI Truly Become Your Work Partner | BlueRun Family Headlines
Kimi K2.6 Released and Open-Sourced: Comprehensive Advances in Code, Long-Horizon Tasks, and Agent Swarm Capabilities | BlueRun Family Headlines
