A Physics-AI Startup Raises RMB 400 Million in Its First Year

In Xinjiang, on the edge of the Gobal Basin, there are few people left.

In a similar place, there are millions of roads in the country that are heavily congested.

But in this 4.5 trillion-dollar freight arterial across the country, heavy-duty drivers work long hours, the driving environment is highly dangerous, and the working population is rapidly ageing.

It's a disruption of the workforce and the daily accidents of fatigue driving.

What's needed here is not a smart-driven utility vehicle, but a super-brain that can drive 100 tonnes of steel giants, navigate accurately on unstructured roads, and complete heavy tasks in extremely dangerous environments.

This is exactly what eBay is doing.

They were founded in less than a year with $400 million in funding, but they didn't choose to kill themselves on crowded passenger car racetracks or tell stories about humanoid robots.

As a physical AI company, eBay has come up with a highly disruptive solution:A foundation model of physical AI for reloading scenarios.

How can a physical AI modeling company build a larger modeling capability with a product that has real needs and commercial pathways?

This question is before all physical AI entrepreneurs, and the answer given by eBay is:It's heavy.

Heavycards are not only the commercial launching point of eBay's recent commercialization, but also its physical AI model into the real world, completing its self-evolution as a "supercarrier".

Many people will wonder: is it possible to move a model of a passenger car running directly to a heavy-duty truck with intelligent driving technology that is so advanced today?

The answer is:Absolutely not. The laws of physics will teach you how to be a man.

A passenger car weighs about 2 tons, while a loaded truck can easily exceed tens of tons or even hundreds of tons. Both differ in physical inertia, brake distance and dynamic constraints.

If you put a passenger car's AI model in a heavy-duty "brain", it will behave like an "alien" that tries to get stuck at high speeds like a pickup truck, rushed, completely ignorant of the laws of physics, and the consequences are unimaginable.

"To train a physical AI to overload a scene, you have to have real interactive data specific to that scene".

The founding team of eBay is keenly aware of this.

For them, real AI in physics is not just about understanding the world, but more importantly, understanding physics.

He needs to be able to accurately predict what the consequences of each action will be in the real world, as well as be able to complete tasks under the constraints of complex overload mechanics.

This explains why eBay is going to"Heavy card".It was the first commercial carrier.

It provides a perfect, high-frequency, continuous physical interaction scene.

It runs on real roads every day, dealing with dust, unstructured roads, complex traffic flows and the difficulty of controlling large bodies.

Heavycards are not only a product that has recently become commercially viable, but also the best carrier to perceive the physical world, get feedback, and continuously train "heavy physical AI brains".

Currently, they have operated around 100 pure-electric heavy-duty trucks and 60 routes in 12 regions.

In order to meet the challenges of the real physical world, eBay has built a set called the World Wide Web."Efficient Physical AI"The two-tiered technical architecture.

This architecture cleverly distinguishes the "repeatable universal underpinnings" from "special capabilities for specific tasks" and sets the stage for future cross-device generalization.

Level 1: Understanding the Physical World

Imagine if you were going to teach a robot how to work, you had to teach it what space is, what gravity is, what obstacles are.

It was created by eBay.OmniBase (Multimodal Pre-Training Base)andSwiftWorld is an efficient physical native world model.。

Through a large amount of multimodal data, models learn how to reconstruct multimodal spaces, understand geometric relationships, and predict the dynamic laws of the future of objects.

At this stage, AI learns the "common sense" of the entire physical world.

Level Two: Special Ability to Become an Old Driver

It's not enough to have common knowledge, but it's also necessary to have sophisticated skills.TwinFast is a dual-speed synchronous cognitive driving system.andL4Forge (after reinforcement training)。

The TwinFast architecture combines "logical reasoning" and "low end-to-end latency" to accommodate complex scenario judgments and milliseconds of maximum response.

At the same time, through L4Forge reinforcement learning, the rules of traffic, the safety bottom line -- such as never tracking down, you have to keep a safe distance -- are engraved into the DNA of AI.

Behind it is a gambling company that accounts for about 76% of the "All-Star Squadron", with 170 people with experience in "front-end model research and development" and "hardcore L4 engineering mass production deployment".

For example, CTO is a core member of the first-generation Apollo, who has led the pre-packaging production research and development of more than 4,000 heavy-duty trucks and has over 500 million kilometers of commercial operations support experience, and is best known for how to make AI "live on heavy-duty trucks".

In addition, the team is responsible for the end-to-end integration of Apple/Bosch, some lead the ADS end-to-end production of Huawei, some build the ideal data closure, and others master the all-in-one training framework and so on.

In the era of big models, the data maker has to get down to earth.

However, real-world real-valued data is extremely scarce and expensive for a physically overloaded AI. You can't buy it, and you can't rely entirely on simulation synthesis.

Eibot's approach was very "wary".

They have not closed their doors in the lab, but have been entering the real production line since Day 1. Currently, eBay has deployed about 100 pure electric heavy duty cards in 12 regions such as Siemens, Xinjiang, and Washington, covering 60 real business transportation lines.

In other words, eBay is producing commercial value for real gold and silver every day while other companies are struggling to find a test scene.

At the same time, these about 100 cars are like tireless "data collectors" who are madly absorbing first-hand experience from the physical world.

A set of data is enough to explain the rotational speed of the propeller.

New data per month:More than two million clips.

Added automatic tagging:It's over 100 million tons.

It's all about ROI technology, and it's a mess.

Why didn't the last generation of L4 do what eBay did?

Applications of autonomous driving on heavy cards are not new, but looking back at the history of L4 heavy card players, we find an embarrassing reality: they are mostly trapped in the mud of "technical self-esteem".

Scenic misplacement, lack of real operational data, purely technical deliveries and no operational underpinnings... these historical bottlenecks have left the last generation of players unable to run through the business cycle.

Physical AI has become extremely widespread and can only occur in real physical enclosures.

Eiboto has chosen "0-300 km of large-scale shortfall" (from mining areas to coal-washing plants, ports to distribution centers) as the first commercial landing point.

It is a market for road freight of up to 4.5 trillion dollars and a "pain point" market that is desperate for AI to save.

Keep in mind that self-driving passenger cars are about "social values" and "safer travel".But unlike HeavyCards, HeavyCards are purely a "production tool" and logistics companies only look at a little bit of hydrogen that can't reduce the added cost.

Currently, there are about 8 million heavy-duty road cards in the country, and the market for road freight is worth up to 45 trillion dollars. In this huge market, there is a deadly pain: the cost of driving is extremely high and the labor force is severely fragmented.

Heavy-duty truck drivers work in extremely harsh environments, often eating in their cars, driving fatigue and overloading are often the cause of accidents. Today, the average age of the industry is approaching 50, and young people in their 80s and 90s are not willing to do this at all.

In heavy-duty TCO (full-life-cycle costs), the driver's manpower costs account for up to 30%-35% and have even exceeded the electricity costs of new energy heavy-duty charging, becoming the largest single cost.

The commercial logic of eBay is very simple and rough: to "liberate" drivers from the cockpit. This is a market of up to 1.2 trillion dollars.

"What we're doing is pragmatically getting the driver off this dangerous, high-intensity car and saving businesses one-third of the cost".

In order to ensure the technology can land at the fastest speed, eBay did not choose the most complex open high-speed lines, but precisely cut into "great 0-300km short transfers" (such as coal, sandstone, chemicals, etc.).

The roads are relatively controllable, and many of them are mines to coal-fired power plants, and there are very few social vehicles.

How did you get the money?eBay has designed a "two-step" business closure:

Stage one: TaaS: Transportation is a service.By financing the lease, eBay has already invested about 100 pure-electric heavy-duty cards for self-driving car fleet transportation. Taking transportation charges on real business lines, running from data collection, model conversion to customer delivery, demonstrates the safety and economics of AI drivers.

Stage two: DaaSAfter successful verification, eBay will switch to the "Light Asset" mode where customers (or owners) own vehicles, and eBay will only provide "AI Driver Technical Services".

IBM is the only truly manned ground force in the industry that has managed 1000-degree unmanned equipment, delivering real transport productivity on-site to customers since Day 1.

If you think that eBay is just a company that makes "driving trucks", you're underestimating it.

It's hard to believe that the first "milling stone" in the real world was taken from a large model of physical AI.

In real heavy industry scenes, heavy cards are often at the center of the entire chain of operations. Its starting point is the excavator and the loader, and the end point is full of imagination.

Eibot's long-term strategy is to use the general physical world representation trained in the first stage of the heavy-duty world to generate more heavy-duty devices horizontally!

"Expansion across the body".It's the sexiest word in eBay's tech architecture.

Since the model already knows how to move, avoid obstacles, and understand kinetics in this 3D world, the brain can adapt quickly to other devices by switching to a targeted "operating action library".

In fact, eBay has started collaborating with leading research institutions to use first-person data to study the physical unloading of excavators and loaders. In the future, in the same mining area or port:

The entire overloaded industrial pipeline will be coordinated by eBay's "same AI brain".

What eBay wants to do is real industrial-grade AI, real productivity-free AI that promotes productivity development.

Today, there are many stories of AI creativity and sexyness, but the real thing is that the key is to make it work hard and save costs.

The answer on eBay is very convincing.

From having a real need to be able to calculate the economic accounts; to accumulating the richest overload data in the entire industry with real deployments; to harnessing high-efficiency, widespread physical AI architectures through "war rehearsal"; and finally to bringing the capabilities to the wider heavy-industrial landscape.

In an age of aging drivers and a severe shortage of heavy-duty labor, eBay is using a unified physical AI brain to connect transportation and unloading into a continuous business cycle.

This is the hard core path to the sexiest, most physically intelligent thing.

By 2027, eBay is expected to reach both ends of low-speed heavy-duty trucks, launching mass commercialization; launch the world's fastest-landing L4 heavy-duty truck fleet on public roads in 2028; and expand to more heavy-duty industrial equipment in 2030, hitting the billion-dollar market.

The second half of the unmanned aerial vehicle, the stellar ocean of physical AI, has just opened.