Gasgoo Munich- "Last year, discussions on embodied intelligence revolved around demo videos. This year, that's no longer enough. We have to get into real-world scenarios and look at actual operations and specific task completion."
A recent remark by an industry insider at WAIC 2026 struck at the heart of the current pivot in embodied intelligence. As bipedal walking and dexterous manipulation are no longer scarce technologies, the sector's benchmark has finally shifted from a "tech showcase" to an "industrial proving ground."
Yet, contrary to the popular image of robots serving coffee or folding clothes in domestic settings, the first to accept this challenge isn't home services or commercial reception. It is logistics and warehousing—a seemingly traditional sector.
Underway: From Tech Demos to Testing
In May, U.S.-based Figure publicly livestreamed a 200-hour session featuring a humanoid robot sorting packages.
Image Source: Screenshot from Figure livestream
In the footage, the Figure 03 robot stood by a conveyor belt, autonomously executing processes such as scanning, grasping, flipping, and depositing packages. Over 200 hours, it processed approximately 249,600 parcels—sorting nearly 21 items per minute on average.
Figure's livestream serves as a microcosm of embodied robots entering the logistics and warehousing space.
As embodied intelligence rapidly moves from the lab into real-world scenarios, many players are treating logistics and warehousing as the "first stop" for pilot deployments.
ROBOTERA, for instance. The company has built a full-stack embodied intelligence solution for feeding small logistics items. It has already partnered with SF Express and China Post to achieve normalized operations at over a dozen logistics centers nationwide.
Powered by its full-size half-body humanoid robot, the M7, along with its proprietary direct-drive five-finger dexterous hand, the XHAND1, and the ERA-42 embodied brain, the solution can handle packages of various sizes, shapes, and materials. It has previously achieved a feeding speed of up to 1,200 items per hour, reaching over 85% of human efficiency.
Image Source: Star Era
As a half-body modular humanoid robot optimized specifically for logistics sorting stations, the M7 features a "half-body + column" design. The logic is simple: sorting is a typical fixed-station task where the robot doesn't need wide-ranging movement. A fixed form maximizes operational stability while significantly lowering hardware costs and extending continuous runtime.
In Figure's aforementioned 200-hour sorting test, aside from battery swaps during shift changes, the Figure 03 barely moved from its spot.
However, while the M7 lacks autonomous walking capabilities, its 1,660 mm arm span and 3-degree-of-freedom waist design allow it to operate over a 2.1-meter diameter range—fully meeting the demands of a single sorting station.
Unlike the Star Era M7's focus on fixed stations, the Elf G2 Max features a wheeled humanoid design. Combining a dual-armed humanoid upper body with an omnidirectional wheeled chassis equipped with a lifting mechanism, it handles full-case palletizing and tote transfers during inbound operations, bridging the "last mile" of logistics flow from unloading to storage.
Geek+, a leader in warehouse logistics robotics, went so far as to establish a dedicated subsidiary for embodied intelligence last year. Targeting B2B scenarios like logistics and manufacturing, it develops technologies and businesses for robotic arm picking and general-purpose robots. In February of this year, it officially released its first general-purpose robot for warehousing, the Gino 1.
The Gino 1's "embodied brain," Geek+Brain, is deeply fused with the massive warehousing data Geek+ has accumulated over the years. It also adopts a wheeled chassis design similar to Zhiyuan's Elf G2 Max, capable of handling multi-tasks such as picking, box moving, packing, and inspection. Mass production is slated to begin gradually in the third quarter of this year.
Other wheeled robots, such as Galbot's S1, Fourier's GRW, Leju's KUAVO 5-W, and MagicLab's MagicBot D1, have also identified logistics and warehousing as a key deployment scenario.
This makes it clear: despite the hype surrounding humanoid robots, the industry is largely opting for a more pragmatic "wheeled, stripped-down" approach for logistics deployments. This balances the efficient mobility of a wheeled chassis with the dexterous manipulation capabilities of a humanoid form.
Image Source: JD Logistics
Driven by the collective efforts of the upstream and downstream supply chain, the deployment of embodied intelligence in logistics is gradually moving beyond the laboratory demo phase. It is now entering a period of normalized verification in real business scenarios and small-scale commercialization.
Why is Logistics Leading the Charge?
Logistics has become the first stop for embodied robot pilots not as an accidental choice driven by trends, but because its scene attributes align perfectly with the current stage of the technology.
First, from a demand perspective, the logistics sector has a massive need for labor and faces prominent structural labor shortages. Consequently, there is a high willingness to pay for embodied robots.
In logistics, roles like express sorting, warehouse picking, and palletizing are characterized by high intensity and repetition. Night shifts and heavy-lifting roles, in particular, face a perennial dilemma of being hard to staff and even harder to retain. For companies, the demand for robot replacement is not just about "cutting costs and boosting efficiency"—it is an "operational necessity."
Image Source: Leju Robotics
Because of this, it is evident that beyond embodied intelligence startups, logistics and e-commerce giants like JD.com and SF Express are also highly proactive in deploying embodied robots in their operations.
JD.com, in particular, has not only opened its warehouses as testing grounds but is also advancing its own R&D. It has incubated a "Wolf Tribe" of embodied robots covering the entire logistics chain, including storage, transport, sorting, and last-mile delivery.
Secondly, from a scenario perspective, the massive volume of logistics and the reusability of skills provide solid support for the scaled deployment of embodied intelligence.
"The logistics sector is massive in scale and can accommodate a vast number of different robots. Furthermore, the skills we extract from logistics scenarios—such as the simple task of moving an object from point A to point B—can be reused across various scenarios. This means marginal costs will decrease while marginal benefits rise, allowing costs to be continuously amortized through scale," notes Tian Ye, founder and CEO of RoboScience.
According to analysis by Roland Berger, the logistics robot market will reach approximately 70 billion yuan in 2025, primarily applied to internal logistics in factories and warehouses.
Although embodied intelligence products currently account for a small share of this market, their growth trajectory is highly certain. As demand for e-commerce, instant retail, and flexible manufacturing continues to grow, companies are increasingly seeking warehouse automation and smart transport solutions. Combined with the rigid requirements for 24/7 operation and high turnover efficiency in logistics, this sector will be one of the easiest for robots to generate a return on investment at scale.
In the view of a Geek+ executive, the logistics sector encompasses millions of SKUs, covering almost all common items found in daily life. This implies that the stable grasping capabilities robots hone in warehouses could become "atomic capabilities"—fundamental building blocks—for applying robots to other commercial and even domestic settings in the future.
Image Source: Geek+
Furthermore, the characteristics of logistics—"semi-structured environments plus clear task boundaries"—are a perfect match for the current technical capabilities of embodied intelligence.
According to Liu Lige, head of warehousing embodied robots at JD Logistics, a scenario suitable for large-scale embodied intelligence deployment should meet the "three highs and one low" standard: high frequency, high standardization, high quantifiability, and low cost of failure.
High frequency means the scenario must have sufficient demand and enough human labor to justify robot participation. Large business volumes help amortize construction costs and generate enough data to drive a data feedback loop. High standardization implies the environment is relatively standard and controllable. High quantifiability means the value created—such as sorting one item or completing one task—can be clearly calculated by the piece, making the ROI transparent. Finally, low cost of failure means that even if a task goes wrong, it won't result in significant loss or safety risks.
"Regarding specific scenarios, I believe warehouse picking is an excellent direction. It meets the 'three highs and one low' standard and allows us to iterate algorithms relying on massive, diverse SKU data," says Liu.
For instance, as a typical semi-structured environment, warehouse logistics has relatively standard operating environments and workflows. In his view, this falls squarely within the comfort zone of current embodied intelligence capabilities. "Moreover, in a warehouse setting, if a package is dropped, you just pick it up. The scenario is recoverable and doesn't cross safety red lines—which is fundamentally different from scenarios with extremely low tolerance for error."
Tian Ye agrees: "The logistics sector presents high difficulty requirements without exceeding our current capabilities. At the same time, it brings a continuous stream of data back—especially data from failure cases—which helps improve the model and the system's overall capabilities."
Moreover, in the long run, the natural globalization of logistics opens up further long-term possibilities for embodied intelligence.
"Logistics is inherently an excellent international scenario. In many overseas countries, the share of manufacturing is no longer high, but logistics always follows population and demand patterns. The essence of the industry remains labor substitution, so internationalization will be a key theme for future industry development," adds Luo Tianqi.
This is particularly true in European and U.S. markets, where rising labor costs and recruitment difficulties are becoming increasingly acute. Companies are desperate for automation solutions. Highly standardized tasks like inbound processing, sorting, and palletizing provide an ideal testing ground for the current generation of embodied intelligence robots to go global.
According to forecasts by Interact Analysis, by 2030, hourly wage growth in mature markets like the UK, Sweden, and Australia is expected to reach 16%–22%. In some emerging markets like the Czech Republic and Turkey, increases could even exceed 100%.
This implies that deploying embodied intelligence robots in logistics and warehousing is not only realistically feasible but also inherently possesses the potential for global replication.
The Eve of Scaling: Three Unavoidable Barriers
While logistics is widely recognized as the starting track for embodied robots, achieving true large-scale commercialization is far from easy. Currently, to move from pilot deployments of hundreds of units to replication in the thousands or even tens of thousands, the industry must overcome at least three core barriers.
The first barrier is the hard constraint on the lower limit of product capability. This directly determines the life or death of scaling.
"At launch events, people look at the upper limits of a robot's capabilities. But from the perspective of large-scale deployment, we look at the lower limits. Can the robot work stably in the worst lighting and the harshest environments? Can it run stably for three months straight?" says Liu Lige. This is what determines whether a warehouse can ship normally.
Breaking it down further, Liu points out that scaling involves considerations across multiple dimensions: technical metrics, production metrics, and financial metrics.
"The core technical metric is the task completion rate. This doesn't refer to the success rate of a single grasp, but the complete rate of the entire unmanned process." For example, a 95% single-grasp success rate announced at a launch might look impressive on paper. But in a real warehouse with tens of thousands of daily operations, that could translate to thousands of human interventions a day—which is clearly unacceptable.
"Therefore, for logistics scenarios, technical metrics need to reach four nines—a 99.99% task completion rate. We also require that the robot be able to recover autonomously when errors occur," Liu notes.
From a production perspective, he highlights three aspects: first, production efficiency, or whether the robot's UPPH (units produced per person/hour) can match or exceed human levels; second, stability, or whether the robot can work for long periods in harsh environments with an acceptable failure rate; and third, system integration capability. A robot is not an isolated system; in a warehouse, it must interact with WMS (warehouse management systems) and dispatch systems, integrating into the upstream and downstream workflow. It must withstand peak pressures, especially when the gap between order peaks and troughs can be several-fold.
Image Source: Leju Robotics
This means that beyond high efficiency and reliability, the deployment of embodied intelligence in logistics cannot come at the cost of massively overhauling existing warehouse systems, as this would only drive up deployment expenses.
Ultimately, all metrics boil down to finance: can the return on investment cycle meet corporate expectations? After all, no matter how advanced the technology, if the ROI doesn't meet the commercial baseline, true scale deployment will be difficult.
The second barrier is the dilemma between cost and scaling.
Compared to traditional robotic arms and AGVs, hardware costs remain high for both humanoid and wheeled composite embodied robots. Core components like dexterous hands, high-precision sensors, high-performance joints, and specialized computing boards directly drive up the price of a single unit and lengthen the payback period.
Moreover, there is significant room for improvement in the stability of core components like dexterous hands and high-performance joints.
Take dexterous hands as an example. According to a Geek+ executive, the lifespan of dexterous hands from leading companies is currently only 300,000 to 500,000 cycles. A warehouse robot needs to perform 8,000 to 10,000 grasps a day, reaching the limit in just a month. For this reason, Geek+ chose to avoid the dilemma of "two fingers not being enough, five fingers being too fragile" and developed a three-finger dexterous hand, seeking a balance between dexterity and industrial reliability.
Logistics, in turn, is a field extremely sensitive to cost and reliability. If the marginal benefits of embodied intelligence robots fail to cover the comprehensive costs of procurement, deployment, and operations, large-scale adoption will remain elusive.
The third barrier is the shortfall in engineering and the industrial ecosystem.
Many assume that once a robot is developed, it can go straight to work in a warehouse. That is not the case.
In a logistics setting, an embodied robot is not a standalone automation device but a node in the warehouse workflow chain. It must integrate into the enterprise's existing digital ecosystem, such as WMS and dispatch systems. It also needs to adapt to the layout and workflows of different warehouses, and even coordinate with on-site human workers.
This means that beyond robot development, companies must complete significant engineering work involving system integration, process adaptation, and on-site debugging. The industry has yet to establish unified interface standards and operational norms. Adapting to different scenarios mostly requires customization, and the workload and difficulty of this engineering implementation are no less challenging than the R&D breakthroughs of the robot technology itself.
From this perspective, the challenges facing embodied intelligence in scenario deployment are not isolated bottlenecks for a single company. They are the growing pains the sector inevitably experiences as it moves from "technical verification" to "industrial implementation."
In past years, the industry became accustomed to defining technical progress with demo videos and judging sector heat by funding amounts. Now that we have reached the deep waters of implementation, we need to settle down, hone reliability in real business operations, amortize costs, and build up engineering capabilities.
There are no shortcuts in this process. It requires continuous collaboration between upstream and downstream players within real-world scenarios to solidify the roadbed leading to the future of general embodied intelligence.
