Author | Su Jialing
This year, the spotlight on Physical AI has shone more brightly on robots. Embodied robots have attracted the most funding, the most media coverage, and dominated the most bustling exhibitions in the Physical AI track. However, Physical AI is not limited to just one form. If embodied robots address "how machines work in the physical world," then the other path centered around cargo—Cargo Intelligence—addresses a more practical issue: how machines can continuously make money in the physical world.
The narrative around robots is booming, but the first to achieve a closed-loop business model is on the cargo side. Especially in freight scenarios, where transportation demand is rigid, scenarios are relatively closed, costs must be accurately calculated, and the compliance path is clearer.
An L4 autonomous vehicle driving itself from Point A to Point B does not mean a freight business is established. What truly determines commercialization is where a ton of cargo departs, where it ends, how many people, vehicles, time, and energy are required, and ultimately, how much money is made per ton of cargo. KargoBot is attempting to answer this question.
From proposing the "human-driven lead vehicle, unmanned following vehicles" hybrid intelligent platooning in 2021, to achieving a positive economic unmanned freight closed-loop at the logistics route level in Ordos in 2025, and now expanding to core logistics corridors across six provinces, KargoBot is not following the single-track logic of "building more unmanned trucks." It first makes AI freight viable on a single route, then turns this capability into a replicable module, and finally connects these modules into a network—moving from point to line, and from line to network.
What KargoBot truly seeks to scale is not an unmanned truck, but isa replicable, transferable, and networkable Cargo Intelligence system. The true unit of scale for unmanned freight might not be a single vehicle, but a route. And when routes begin to interconnect, the unit of competition shifts from fleets to networks. A self-growing Cargo Intelligence network might be the real fortress worth fighting for after scaling.
01 Calculate the Economics First, Then Talk About Unmanned Operations
In the cargo branch of Physical AI, technological leadership does not equal commercial leadership. The ticket to commercialization belongs to those who first calculate exactly how much it costs to transport a ton of cargo to its destination. The first principle of unmanned freight is not "unmanned," but unit transportation cost. The autonomous driving industry easily treats "unmanned" as the end goal. But freight customers do not pay for L4 itself; what they ultimately buy is the transportation result. Therefore, what Robotruck must pass is not just the technical hurdle, but also the economic account of transportation. According to a rough industry breakdown, traditional freight costs are approximately "33211": 30% for tolls and bridge fees, 30% for energy, 20% for labor, 10% for maintenance, and 10% for depreciation.
The first step is to calculate "labor." L4 can achieve "unmanned" operation, which seemingly means simply cutting labor costs is enough. KargoBot made a decision that seemed like "a step back" at the time: instead of making all vehicles fully unmanned in one leap, they opted for a hybrid platoon, with one human-driven lead vehicle guiding 1 to 5 L4 following vehicles. In fact, the hybrid platoon solves a practical issue that pure technical solutions cannot bypass: in real-world trunk line freight environments, L4 vehicles face far more than just driving itself. Some national highway toll stations lack ETC and require QR code payments, traffic police direct vehicles into weigh stations for load checks, and around mining areas, puddles and dirt piles can appear at any time on the road.
In traditional transportation, one vehicle corresponds to one driver. The hybrid platoon transforms this into one person managing a fleet. The human-vehicle relationship is redefined, and so are the costs. Algorithms control the throttle and brakes more smoothly, reducing rapid acceleration and deceleration, continuously saving fuel and electricity under full-load conditions, and dispatch optimization reduces empty runs and waiting times. Compared to human-driven fleets, the profit margin for transportation can increase by 10%-18%, while bringing a 5%-10% optimization in energy consumption. Ordinary autonomous driving simply adds a system to a traditional truck; Cargo Intelligence, however, recalculates the costs of the entire transportation system. Thus, autonomous driving truly enters the cost structure of transportation companies for the first time. While unmanned operation is the ultimate vision, this step is about reallocating the driver as a production factor within the transportation system, shifting from 1 person to 1 vehicle, to 1 person to N vehicles.
Over the past few years, new energy vehicles have driven down energy costs, but labor costs have been rising. Five years ago, recruiting drivers in Ordos cost just over 10,000 RMB, and it was relatively easy to find them. Now, it is hard to recruit, and turnover is high; there are almost no post-95s in the cab, and night shifts are particularly short-staffed. A person's daily working time is 8-10 hours. KargoBot chose a more pragmatic path: first prove the economics of the platoon, then gradually increase the unmanned rate. Besides the hybrid platoon, KargoBot has two other product lines: L4 single-vehicle autonomous driving and transport robots, promoted in closed and semi-closed scenarios, with the economic calculations delving deeper layer by layer.
The second step is to calculate the vehicle's driving costs. KargoBot's L4 single-vehicle intelligence solution uses single-vehicle L4 capabilities to complete unmanned transportation, increasing the profit margin to over 20%. The third step is to redesign the vehicle itself. The cab-less transport robot solution, KargoBot Space 2.0, removes the cab, increasing cargo space by about 25% and effective payload by about 10%. It supports 7x24 hours of uninterrupted operation, increasing the economic benefit per vehicle by about 30%. Heavy truck air conditioning has always been a major power consumer; after going unmanned, this significant energy expenditure is also saved.
What KargoBot truly changes is the freight cost structure itself. From one-person-multi-vehicle to single-vehicle unmanned, and then to eliminating the cab, it calculates labor first, then the vehicle, and finally the cost of the entire transportation system. However, the same route can also be a completely different business. Different vehicle models, different energy sources, and a ten-percentage-point difference in the proportion of autonomous-driving-available road sections lead to different optimal solution combinations and unit economics. Through long-term delivery, KargoBot has gradually internalized a TCO calculation tool. It incorporates fleet size, vehicle models, energy types, the proportion of autonomous-driving-available road sections, mileage per vehicle, obstacle scenarios, as well as depreciation, labor, maintenance, and energy into the calculations. This means that what autonomous driving companies truly deliver to customers is shifting from a set of technologies to a transportation solution viable in specific scenarios.
KargoBot has achieved a positive economic unmanned freight closed-loop in Ordos, realizing route-level UE (Unit Economics) turning positive, and completing scaled unmanned freight door-to-door operations. And a route that makes money is not just about "making money"; it is the prerequisite for network effects to hold. Without positive unit economics at a single point, all narratives about network effects are castles in the air.
02 The First Route Makes Money, the Second Route Can Be Replicated
A route with positive UE can prove the business model works. But scaling follows a different set of rules. The second hurdle for Robotruck is ensuring the next route doesn't have to start from scratch.
The true value of a route lies not only in its own profitability but also in showing the enterprise "how the next route can make money." At this point, data becomes crucial. But the data here is not just the autonomous driving data commonly understood. The world heavy trucks face is more complex than that of passenger cars. On one hand, the dozens of tons of payload, the ten-plus-meter trailers, and a whole set of physical constraints make empty and fully loaded trucks almost two different vehicles. When fully loaded, inertia is multiplied; at the same speed, the braking distance is significantly lengthened. On the other hand, single vehicles and high-speed platoons are two different systems.
The most extreme physical challenges of platooning are exposed only at second-level following distances: a very small trajectory change by the lead vehicle can be progressively amplified by the following vehicles, creating an "accordion effect." Once a deviation occurs in far-end trajectory prediction, forced corrections can, in extreme cases, cause trailer swing or jackknifing. Therefore, between knowing how to run one route and knowing how to run all routes lies a massive amount of physical scenarios. Simulation and passenger car data cannot replace these real physical samples.
Data for Cargo Intelligence can only grow in real trunk lines and heavy-load conditions. The reason KargoBot started in Ordos is due to both scenario considerations and data inevitability. The transportation scenarios in the northern energy belt are highly isomorphic to those in Inner Mongolia: highways, national roads, provincial roads, village roads, dirt roads, and mining area roads almost reappear in every new province. In other words, Ordos is not an isolated showroom but a microcosm of trunk line logistics scenarios across northern China.
KargoBot L4 platoon identifying cattle and sheep herds. KargoBot here has polished complex road conditions, unstructured environments, and real freight operations. After this, expanding outward provides a scenario library for reference and replication. More importantly, only when the operational frequency of freight is high enough does AI have the opportunity to continuously encounter problems. On a route from Guangzhou to Suzhou, the top logistics companies only dispatch 2-3 trucks a year, with a total of 6 trucks on one route.
In Inner Mongolia, from a coal mine to a power plant, there are 400 trips a day, and on some routes, even 4,000 trips a day—the magnitude is completely different. KargoBot's structured benchmark dataset NuScenes-S, covering scenario descriptions, object perception, and driving decisions. Every single freight transport leaves behind real data. Currently, KargoBot's fleet of over a hundred L4 autonomous vehicles and tens of millions of kilometers of platooning autonomous driving mileage have accumulated over 4 million edge scenarios in trunk line freight, including road debris, cattle and sheep crossing, strong crosswinds, long slopes, and changing road surface adhesion conditions. No matter how large the scenario library is, it is merely a static asset.
What truly brings down the marginal cost of the "next route" is a cyclical data production chain: real transportation exposes problems, problems become scenario data, data trains models, models return to the vehicles, and then enter the next real transportation run.The starting point of data is technology, and the endpoint is the economic ledger. This is the true sense of scale effect in unmanned freight: for every new route opened, there are fewer pitfalls to navigate from scratch. The value of data also shifts from training models to further reducing replication costs. Its greater significance lies in ensuring the second route doesn't have to start from zero. Scaling begins here.
03 "Weaving Lines into a Network": Beyond Fleets, the Competition Shifts to Networks
Scaling cannot be simply understood as fleet size; two concepts need to be distinguished here. 100 unmanned trucks can be a very large fleet, but if they are independent and run separately, they are just 100 units of fixed assets. The scale effect of a fleet is linear, it is addition. But a network is multiplication, with a stricter definition: adding a new node can increase the value of other nodes, nodes share resources, leveraging the entire network. A customer's freight scenario can train the entire model, making all vehicles smarter together; a car manufacturer joining brings more transport capacity; a safety system replicated to more regions promotes standards to every kilometer.
When 100 vehicles enter the same network, each newly added node can potentially improve the utilization rate, data volume, and dispatch efficiency of existing nodes. A route may correspond to one customer and one project. When routes become a network, it becomes "one mature template + a set of standard capabilities + a new scenario." KargoBot summarizes this path as "1+N+X".
"1" is not just a simple template route; it validates a complete transportation closed-loop: technology, customers, vehicles, energy, stations, operations, and economics. "N" means cross-regional expansion, replicating mature solutions to more regions and routes. KargoBot is extending its network to core logistics corridors in Inner Mongolia, Ningxia, Gansu, Xinjiang, Shaanxi, and Shanxi. "X" continues to add cargo types and scenarios on already formed nodes.
Expanding from bulk commodities to express delivery, industrial manufactured goods, components, agricultural and livestock products, as well as ports, railway multimodal transport hubs, and logistics parks, serving more transportation tasks. This is actually growth in two different directions: N is horizontal expansion, spreading the network; X is vertical replication, weaving the network denser. It is reported that the average deployment cycle for KargoBot's new routes has been compressed from 8 months to 3-4 months. The capability that truly connects various routes comes from three levels. First is hardware standardization. The physical carrier of Cargo Intelligence is becoming platformized. KargoPlatform Gen5.0 supports multiple OEMs; central computing + zonal controllers reduce electrical complexity by 50%, eliminating the need to redevelop an autonomous driving system every time a new truck manufacturer is involved.
Second is software platformization. KargoCloud integrates vehicles, cargo, transport capacity, energy, and stations into a single dispatch system. When a vehicle enters the network, it is immediately taken over, transforming single-point capabilities into system capabilities.
KargoBot's self-developed KargoCloud dispatch platform. Third is operational template standardization. What is truly hard to replicate is often the experience hidden in projects: safety systems, O&M processes, energy solutions, and TCO tools. Transferring them as standard modules means making these implicit capabilities explicit. At this point, what KargoBot is truly replicating has expanded from unmanned truck driving technology to a comprehensive unmanned freight solution.
When adding a new route becomes more like replication than starting a new business, Cargo Intelligence truly moves from a "project-based" model to a "networked" one. And the true value of a network lies in the connections between nodes. Simply being an autonomous driving technology company means mastering algorithms is enough. But a Cargo Intelligence network is a different matter: vehicles, cargo, energy, stations, orders, and AI must be connected. KargoBot chooses to enter this system together with truck manufacturers, energy companies, customers, and logistics partners.
Truck manufacturers provide vehicles, energy companies own battery swap systems, logistics companies have cargo sources, and customers have scenarios. KargoBot's ARS strategy (AI-ROBOT-SOLUTION) links autonomous driving, dispatch, and operational capabilities. What it ultimately holds is a complete set of network rules for vehicle network integration, dispatch, data generation, and data feeding back into overall network efficiency.
Over the past few years, autonomous trucks have experienced the "hundred schools of thought" era of debates between L2 and L4, and platooning versus single vehicles. As technology moves from scenario demonstrations to the scaling stage, what determines scale is not who owns more isolated unmanned trucks, but who can make transportation routes synergize. From single vehicles to fleets, then to routes, and ultimately to networks, connecting more routes at lower costs. Newly added customers, new cargo types, new routes, and newly connected vehicles all become new nodes in this network.
When one route can make money and the second route can be replicated, a network truly begins to grow. The industry is entering a "winner takes all" stage. The competition is not about larger fleets, but a self-growing transportation network. The future competition is about the infrastructure efficiency of the Cargo Intelligence network.