If We’re Not Building Humanoid Robots, We’re Going All In on Restaurant Back Kitchens
Embodied intelligence startup AtomBite.AI has recently closed a seed round in the tens of millions of RMB, led by InnoTorch Capital, with participation from the Tsinghua Alumni Seed Fund and several well-known individual investors.
The funds will be used primarily to develop embodied world models for restaurant scenarios and to refine and commercialize its core products.
At the same time, the company has also secured deployment interest for product collaboration from several leading companies in China and overseas.
A Startup Built on a Former Meituan Waimai Executive’s Track Record
The company’s founder, Dr. Wang Dong, previously served as the technical lead for Meituan Waimai’s business unit. He managed a product and R&D team of over 1,000 people and led the development of the delivery algorithms, data systems, and architecture supporting tens of millions of daily orders.
During his time at Meituan, he also developed multiple smart hardware products aimed at merchants’ pain points and successfully deployed them.
His academic foundation is equally solid. A student of Academician Zhang Bo at Tsinghua University, he focused on AI representation during his doctoral studies and has continuously published papers at top AI and robotics conferences.
In 2011, he built the world’s first commercial video-based face recognition and tracking system. Later, he went on to create an industry-leading local services recommendation system and a mobile-native advertising platform.
Co-founder Li Tao is also a Meituan veteran who once led Meituan Waimai’s algorithm and data systems.
Another co-founder, Li Haozhe, is a serial entrepreneur with years of experience in global commercialization.
Other core team members come from Tsinghua University, the University of Science and Technology of China, Meituan, Horizon Robotics, and other leading institutions and companies.
A well-known investor commented:
Dr. Wang Dong is absolutely the kind of founder best suited to building embodied applications for commercial scenarios — the combination of “technical depth + global perspective + firsthand operational insight” is almost impossible to find a second example of in this target market.
Not Building Humanoid Robots, Going All In on Restaurant Back Kitchens
In 2026, the embodied intelligence sector is exploding, with one round of funding after another going to general-purpose humanoid robots.
But AtomBite.AI chose a path that may seem less “sexy” at first glance: restaurant back kitchens.
After leaving Meituan, Wang Dong spent extensive time visiting restaurant merchants and researching delivery platforms. He ultimately reached a conclusion: the restaurant back kitchen is the most commercially certain entry point for embodied intelligence.
The reason is straightforward: the demand is globally shared, the pain points are acute enough, and the business model is verifiable.
Specifically:
Global food delivery orders continue to grow, but the handoff from merchant preparation to courier pickup remains a weak link in technological transformation.
High order error rates and spill rates mean users, merchants, couriers, and platforms all end up paying the price.
On the other side, labor shortages in the global restaurant industry have become structural.
Wang Dong’s view is direct:
Merchants do not care whether your robot looks human. What they care about is: how much it costs, what it can do, and what costs it saves me.
As a commercial company, of course we care about the grand vision of general-purpose models, but what we care even more about is the feasible path to a physical model that integrates the digital and physical worlds.
Embodied World Models: Teaching Robots to “Think Before They Move”
So what exactly is the “embodied world model” that AtomBite.AI is building?
At the beginning of 2026, the embodied intelligence field is undergoing a paradigm shift. In the past, robot control was essentially “do whatever you see” — VLA (Vision-Language-Action models) directly mapped the current scene and instructions into actions, much like a conditioned reflex.
The problem is that real back-kitchen environments are extremely complex: oil fumes, occlusions, multi-person collaboration, randomly placed items… With a pure reflex-based policy, generalization is poor, and even a slight change in scenario can easily cause errors.
The World Action Model (WAM) offers a different path: before a robot acts, it first rehearses what will happen next.
More specifically, WAM is trained on large-scale video data so the model learns to predict, “If I take this action, what will the world look like?”
Turing Award winner Yann LeCun has made it clear: only with the ability to “predict the future” can AI perform complex planning.
What AtomBite.AI is doing is bringing this “think before you move” capability into the concrete setting of restaurant back kitchens.
From day one, AtomBite.AI’s embodied world model has been anchored in real commercial environments: What are the physical constraints of a back kitchen? Where are the tolerance boundaries for packaging actions? How should the scheduling logic for multiple devices be structured?
These are questions that simulation alone cannot answer.
A Three-Layer Technical Architecture, from “Able to Move” to “Able to Think”
Based on this approach, AtomBite.AI’s technology stack is divided into three layers:
- The top layer is the embodied world model — enabling robots to form a complete understanding of the back-kitchen physical environment, predict the consequences of actions, and make autonomous decisions. The core is “understanding a scenario”;
- The middle layer is a dual-engine system for task orchestration and scheduling optimization — translating cognitive decisions from the top layer into precise physical execution plans while uniformly coordinating multiple back-kitchen devices;
- The bottom layer is the deep integration of self-developed core components with general-purpose hardware platforms — ensuring the system can reliably complete a wide range of complex operations in real commercial back kitchens, meeting requirements for precision, long-horizon execution, and generalization.
The WAM deployment path starts with takeaway packaging and handoff, tackling the most error-prone part of fulfillment; then expands to controlling existing back-kitchen equipment such as stir-fry machines and dishwashers; the ultimate goal is to build a digital-twin back-kitchen operating system that enables globally optimized intelligent scheduling.
The core logic of this approach is to begin with real-world deployment, then gradually build and validate an embodied world model that conforms to physical laws through actual scenarios, allowing the scenario itself to drive model growth.
The First Year of Embodied Applications Starts in the Back Kitchen
Dr. Wang Dong said:
2026 is the first year of embodied applications. Only by creating a “data-model-hardware” flywheel in real environments can embodied intelligence move from the lab into industry. From day one, we have been building products and a business system for the global market.
A seed round in the tens of millions of RMB is only the beginning for a company that has been founded for less than two months.