As AI enters the physical world, why has JD Logistics chosen an "unsexy" path?
In the 1990s, Denver International Airport once aimed to build a futuristic airport.
The most ambitious part of their plan was to completely eliminate manual handling of luggage: starting from the check-in counter, automatic shuttles traveled at high speed along underground tracks to deliver luggage directly to the corresponding flights.
However, the future did not arrive as scheduled, and instead turned into a nightmare for the airport.
Luggage was misrouted, shuttles jammed the tracks, and parcels were crushed and damaged. The airport's opening was delayed by a full 16 months, with more than 1 million US dollars spent daily on interest and maintenance fees for an idling airport.
Nearly 30 years later, the same story repeated at Tesla's Fremont Factory.
In 2018, the Tesla team designed a complex robot equipped with a vision system to place a soft fiberglass mat on top of the battery pack, but it often failed to grasp the mat or placed it in the wrong position, dragging down the entire production line. Elon Musk then raised a more fundamental question: was this mat even necessary? Comparative tests found that the in-vehicle noise level showed no difference whether the mat was installed or not.
The two stories separated by nearly 30 years share the same core: when the evolution of a single unit or single thread cannot keep pace with the entire system, even the most advanced technology will become a burden. The more capable the machine is, the more efficient production will not necessarily be.
Today, similar fanaticism is recurring in the field of embodied intelligence. According to data from IT Juzi, in the first half of 2026, there were 322 financing deals in China's embodied intelligence sector, with a total amount of about 935 billion yuan, 5 times that of the same period last year. However, the most popular stories in the robotics industry mostly take place in laboratories and demos. In strictly controlled stable environments, robots can fold clothes, cook, and work in factories, and their progress seems very fast.
Few people have calculated this account: how big is the gap between the efficiency and reliability requirements of demo scenarios, household scenarios and industrial-grade scenarios? When the task base increases from dozens of times to 550 million times a day, and when the environment changes from a booth with stable lighting to a warehouse with tens of millions of SKUs, thousands of different parcel shapes and people walking around at any time, how much persuasiveness do those carefully choreographed demos have left?
How far is physical AI from the "singularity" of real implementation?
Logistics: The Training Ground That Pulls Physical AI Back to Reality
At the 2026 World Robot Conference, parcel sorting and tote handling have almost become the most crowded exhibition tracks for embodied intelligence. More than a dozen humanoid robots sort express parcels in perfect unison, and many visitors joked that "there are too many robots, not enough express parcels to sort".
The bustle is real, but most of these scenarios are scattered, fragmented and unconnected: a robot completes a perfect sorting action on the booth, which is just a drop in the bucket compared to the astronomical level of system complexity of the real-world logistics industry.
The real logistics environment poses high-density and high-intensity practical tests for physical AI.
Do you know how many express parcels are on the road in China every day? According to data from the State Post Bureau and CCTV Finance, in the first half of 2026, China's average daily express business volume was about 550 million pieces, and during major promotion events such as the 618 Shopping Festival, the maximum daily business volume can reach 777 million pieces. Behind the 550 million parcels are 550 million times of collection, transportation, sorting and delivery. Under high-frequency physical actions, the fault tolerance rate is pushed to the extreme.
At the same time, the complexity of the logistics industry far exceeds people's imagination.
Taking JD Logistics, which integrates warehousing, supply chain and distribution, as an example, according to financial report data, as of June 2026, the total managed area of JD Logistics' warehousing network exceeds 36 million square meters, which is equivalent to about 5,000 standard football fields. More than 70,000 self-operated transport vehicles, 13 self-owned full cargo planes, more than 600,000 self-owned distribution operators, and more than 19,000 distribution stations and outlets are operating on this network.
At the same time, the scenarios it covers are also extremely complex: islands, villages, cities, cross-border overseas markets. The problems to be solved in different situations are completely different, with extremely high non-standardization. Hundreds of thousands of couriers need to rely on manpower and hard work to support this huge system, which is in urgent need of burden reduction from physical AI.
The logistics industry did not start automation in this physical AI boom. As early as more than a decade ago, the logistics industry successively introduced conveyor lines, sorters and handling robots, and gradually completed 80% of the most easily standardized and most easily automated parts.
The remaining 20% is the real "hard bone".
For example, in links such as goods picking and non-standard parcel stacking, the shape of items and the operation environment are constantly changing. If the robotic arm grasps the parcel crookedly, the unmanned vehicle hits an obstacle, or the robot picks the wrong goods, it will immediately affect the subsequent fulfillment links. The warehouse itself rarely maintains an ideal state: there may be hundreds of thousands of SKUs in one warehouse, the placement posture of the same product is not fixed, cartons will deform, soft packages may slip, and temporarily added shelves and floor stacks will change the original moving lines. Order fluctuations further increase the operation difficulty: if robots are configured according to the peak demand, a large number of devices may be idle at ordinary times; if they only cover the daily order volume, manual backup space needs to be reserved for peak periods.
All non-standard problems are superimposed one by one, forming black holes that block the implementation of physical AI and prevent the arrival of the singularity.
The industry is not indifferent. A large number of embodied intelligence enterprises pouring into the logistics industry is a clear example. For example, Starbot announced that it has entered more than ten logistics centers of SF Express and China Post, targeting links such as parcel feeding and sorting; Jiushi Intelligence cooperated with Cainiao to form an unmanned fleet, targeting the transportation link; Sainard's autonomous loading and unloading robot focuses on links such as autonomous unloading, code scanning and palletizing.
Beyond them, JD Logistics also provides an excellent observation sample.
As we all know, due to its long-term directly operated warehousing, distribution and supply chain, the labor, warehouse network and equipment of JD Logistics are directly included in its own cost structure. This means that every 1 percentage point increase in efficiency in each turnover link can theoretically directly affect its own income statement, which is actually an internal ROI verification field with extremely strong driving force.
Not long ago, on September 9th, at the JDD JD Global Discovery Conference, JD Logistics exhibited its self-developed industrial-grade large model "Super Brain 3.0", and intensively released and upgraded a number of new products including Yilang (Heterogeneous Wolf), low-temperature version of Zhilang (Smart Wolf), Canglang (Warehouse Wolf), Mother Wolf Health Version, Dulang (Lone Wolf) 6th Generation Plus, and Feilang (Flying Wolf) L05, upgrading the entire "Wolf Pack" robot system to 9 categories and 11 robots, covering all storage, handling and sorting links in the entire logistics industry chain, which seems to be the perfect answer for physical AI to settle into the real world. But is that really the case?
To understand this chain, we might as well start from the journey of a parcel.
The Journey of a Parcel
Imagine a scenario: you are browsing products in the JD APP, even before you click pay, just dragging the product into the shopping cart, the entire system of JD Logistics has already started running.
"Super Brain 3.0" intervenes first: it predicts orders, analyzes inventory, and plans paths, then sends instructions to every device and every person distributed in warehouses, vehicles and distribution stations, just like the brain sending commands to the nervous system.
If we compare this "Super Brain + Wolf Pack" system to a human, "Super Brain" is its brain, and the "Wolf Pack" robots are its trunk and limbs. The brain is responsible for thinking, predicting orders, planning paths, and scheduling resources; the trunk and limbs are responsible for execution, picking up goods, sorting, stacking, and connecting.
JD Logistics is one of the pioneers in the field of intelligent warehousing in China. As early as 2016, JD Logistics built the first batch of unmanned warehouses in the industry. After that, it gradually tested equipment such as "Dilang (Ground Wolf)" AGVs, "Tianlang (Sky Wolf)" pallet robots, and "Nanny Wolf" sorting walls; in 2023, JD Logistics launched the "Zhilang" goods-to-person system, together with "Feilang" drones, forming the "Wolf Pack" robot legion.
From the underlying naming logic of the "Wolf Pack" robots, we can see that JD Logistics has not emphasized "individual combat" from the very beginning, but imitates the biological characteristics of wolves fighting in groups, so that robots form cluster collaboration in all links of the supply chain. Under the scheduling of "Super Brain", different robots undertake different links of warehousing, picking, handling and distribution, and the value generated by linkage collaboration is far greater than the sum of the work of each robot alone.
After "Super Brain 3.0" issues the instruction, the trunk and limbs respond immediately. Next, this product will be found, taken out, handled, sorted, transported, and finally delivered to consumers.
The first step is to take the product out of the warehouse.
In the past, pickers had to walk 30,000 to 40,000 steps a day in the warehouse, rushing back and forth to find all the products of one order from different storage locations. The emergence of Zhilang changed this. As the "relay runner" in the warehouse, the "Zhilang" robot lifts up and down the three-dimensional shelves to take out the totes, and the ground shuttles are responsible for handling. The two types of equipment are decoupled and perform their own duties. Its biggest capability is to achieve a 12-meter clear height for storage — while most goods-to-person products in the industry can only utilize a clear height of 3 to 5 meters.
Moreover, in terms of "cost accounting", R&D personnel of JD Logistics revealed to 36Kr that the team reversed the unit cost and R&D plan through order volume, rent, floor efficiency and labor efficiency, and compressed the payback period of Zhilang to 2 years. In the capital-intensive industry, a 2-year payback period is almost astonishingly fast. The general requirement for automated equipment in the logistics industry is 3 years, and JD Logistics has directly increased the speed by 1/3. According to JD Logistics' data, the Zhilang system currently supports nearly 1 million orders of fulfillment per day internally, and has been put into operation in overseas warehouses in the UK and Germany.
By the way, there is also a "frost-resistant warrior" in JD's Zhilang family. At the JDD JD Global Discovery Conference on September 9th, the new "low-temperature version of Zhilang" made its debut, mainly targeting cold chain scenarios. In cold storage with a constant temperature of minus 20 degrees Celsius, workers wear cotton-padded clothes to work, their efficiency is less than half of that in normal temperature warehouses, and the turnover rate remains high. The low-temperature version of Zhilang brings tote storage into cold storage: the material is pre-shrunk and compensated, the power supply is replaced with a non-contact wireless solution, and people pick goods in the refrigerated area with efficiency close to that of normal temperature warehouses, completely isolated from the cold storage operation area — people no longer need to enter the freezing warehouse.
But what if the products are stored in ordinary manual shelves, old warehouses, or leased warehouses? This is a larger stock market. The premise of traditional automation transformation is often to "transform the warehouse into a form that robots like", for example, stop production for 6 months, modify shelves, divide aisles, and move storage locations. The vast majority of small, old and aged warehouses cannot afford such costs, and the "cost accounting" is not worthwhile.
This is when "Canglang" comes into play. Canglang is the "picker" in the aisle, it drives directly into the existing warehouse, identifies products in narrow aisles, and stretches its arm to grasp them. Relying on its autonomous mapping capability, Canglang can quickly generate an environment map after driving into the warehouse; it can also take into account the inspection and inventory tasks, and automatically check the goods on the shelves during the gap between picking tasks. There is no need to modify the shelves, no need for the warehouse to adapt to the robot. The robot drives directly into the manually operated warehouse, autonomously picks goods in narrow aisles, and the deployment can be completed in 2 weeks. The coverage rate of the whole warehouse's goods exceeds 85%, the picking accuracy is 99.9%, and the comprehensive picking efficiency is 80 pieces per hour. Don't underestimate the data of 80 pieces. Although it is not yet comparable to the best human workers, it gives the stock warehouses that were forgotten by automation for the first time the qualification to participate in intelligent transformation.
More importantly, cooperating with "Super Brain 3.0", "Canglang" can realize compatibility with the original manual operation. When the production capacity is at the off-peak level, the deployed robots undertake the tasks; when the production peak comes such as promotion events, the system uniformly plans the moving lines of humans and machines at the upstream, and the same SKU can be sorted by both machines and humans, truly bringing physical AI into the warehouse.
After the products are picked and packed, they need to be sorted according to their destinations. We come to the sorting yard of "Yilang".
This is the hardest, most tiring, and most likely link to be "stuck" by machines in the whole chain. In the sorting yard, a skilled stacker has to bend down thousands of times a day to carry parcels of different shapes into cage trolleys. Traditional robotic arms are almost powerless against soft woven bags, bulging air column bags, and irregular special-shaped parts, which is exactly the "non-standard problem" that global automation companies cannot avoid.
At this time, "Yilang" comes into play. Equipped with an industrial dexterous hand, Yilang is the coolest "stacker" on the entire assembly line. R&D personnel told 36Kr that before starting work, Yilang can obtain the order time, outbound warehouse, and destination of the parcel from JD Logistics' internal "brain", and judge the size, material, and distribution direction in real time.
