Embodied Intelligence Enters Factories and Warehouses: What Else Must Robots Compete on After Smarter Brains?

While the efficiency, cost, and commercial prospects of embodied intelligence are still being repeatedly questioned, in factories and warehouses, robots have already undertaken some handling, access, and sorting tasks.

On November 3, the China International Exhibition for Material Handling and Logistics (CeMAT) opened in Shanghai. From climbing shelves and four-way box picking to narrow aisle obstacle avoidance and multi-robot coordination, the technological improvements demonstrated by participating enterprises are refined down to the time taken for a single turn and the occupancy of a single storage slot.

In May this year, Figure conducted a live broadcast of three robots performing parcel sorting for 200 continuous hours; at various industry conferences, many embodied intelligence companies demonstrated operations such as parcel grasping, flipping, and sorting; Xingdong Jiyuan collaborated with China Post and SF Express to deploy relevant robots to logistics centers... Warehousing and logistics have already become an important application scenario for robots.

On October 19, the opinions on accelerating the promotion of technology-driven commerce released by 9 departments including the Ministry of Commerce proposed that embodied intelligence technology should be promoted to land in the commercial and trade fields, expanding application scenarios such as home services and commercial logistics.

However, various types of automated equipment have already been put into use in warehousing and logistics, and newly entering robots must not only adapt to the site but also be compared with existing equipment in terms of efficiency and cost.

The battle of efficiency is as detailed as a single turn

“One is time, one is space, and another is quality.” Wang Zihao, General Manager of the Cainiao Logistics Technology Robot R&D Center, introduced to reporters from The Paper that e-commerce warehouse order flows are large, and some warehouses need to operate 24 hours a day; if a certain product fails to be shipped out in time, it may also affect subsequent links such as joint ordering and transportation. Equipment design must consider efficiency while also trying to minimize the impact of single-point failures on the entire warehouse.

The climbing robot displayed by Cainiao this time can pick up and put down material boxes in four directions and supports double-depth storage, which means placing two layers of material boxes, inner and outer, in the depth direction of the shelf. Wang Zihao explained that before taking out the inner layer material box, the outer layer material box needs to be moved first; four-way picking allows the robot to have more temporary empty spaces to choose from, improving space utilization.

When walking on the ground, traditional right-angle turns require decelerating, rotating, and then accelerating. Wang Zihao introduced that by changing to curved walking, each turn can save about two seconds; as a robot makes multiple turns in the warehouse and multiple devices run continuously, the saved time will accumulate continuously.

According to the introduction, after this upgrade, the climbing robot's operating efficiency is increased by 15%, and the storage density is increased by 20%.

The challenge faced by piece picking is the difference in products. In the same order, there may be both standard products and irregularly shaped, difficult-to-grasp products, posing different requirements for the equipment's processing capability and cost.

Yuanli Lingji this time showcased the first embodied product results launched after the merger with ATOMIX: a multi-agent hierarchical task system for robots. According to the scheme design, about 60% of standard, easy-to-pick items are handled by economical robotic arms, about 20% of complex items are handled by embodied robots, and the remaining about 20% of long-tail tasks are backed up by manual labor and responsible for anomaly handling.

In pallet scenarios, ATOMIX is responsible for upper-level dense storage by four-way vehicles, while autonomous mobile robots handle first-level storage and picking and carrying. According to the description, this “upper storage, lower picking” solution has already been applied in multiple enterprises, with the number of logistics robots in a single project exceeding 100 units.

Original Force Lingji booth

How is the customer's account calculated?

“Every customer's requirements for each robot are different.” XianGong Intelligence Vice President Ding Xia told reporters from The Paper, that if the passage is narrow, the robot needs to be small; if the cargo is large and there is insufficient turning space, the equipment may need to possess lateral movement capabilities.

At the exhibition site, two AI delivery robots with different payloads from XianGong Intelligence were operating in the same area, demonstrating dynamic obstacle avoidance, narrow passage intersection, and S-curve passage; the omnidirectional ground forklift performed autonomous identification and picking of pallets of different types, placement directions, and positions.

According to the introduction, XianGong Intelligence's “robot brain” has already adapted to more than 400 types of core components, supporting more than 2000 robot SKUs. Ding Xia explained that this does not mean 2000 completely different robots; devices with similar appearances, by replacing internal sensors, may also correspond to different SKUs.

The customer's requirements are "fast, good, and economical." Ding Xia stated that the functions must be able to meet complex environments, efficiency must be high, operation must be stable, and the price must also be appropriate; the investment recovery period for equipment in some industrial applications can already be achieved in 1 to 1.5 years.

However, robots also have current capability boundaries. In non-standard abnormal scenarios such as damaged pallets or stretch wrap wrapping, robots may still require manual intervention as a backup, and the experience of skilled forklift drivers still has value. Ding Xia gave an example, saying that factories can hand over some routine transportation to robots while retaining forklift drivers to handle complex tasks—letting machines handle repetitive, stable transportation, and letting humans handle anomalies and experience-based judgments.

Multiple models of embodied intelligence demonstrated by XianGong Intelligence

The person in charge of Yuanli Lingji also told reporters from The Paper that price competition in the domestic logistics robot market is fierce, and profits on some projects are under pressure, but some equipment issues only appear after running for a period of time, and large customers who value quality do not only look at price.

Different customer needs also increased the difficulty of delivery. Ding Xia believes that the cost of changing customer scenarios might be too high, and enterprises need to improve the speed of meeting different needs by adapting components and preparing functions in advance, allowing customers to combine equipment like building blocks, “I standardizing the customization process is also a kind of standardization.”

Cainiao adjusts the upper structure based on standard models according to customer requirements such as carton dimensions. Yang Bin, General Manager of the Logistics Robot Business Department of Cainiao Logistics Technology, stated that standardized products are more convenient for quoting, delivery, and integration, and are also more favorable for cost control.

Yang Bin emphasized that the climbing robot business has just completed the stage from product release to commercial verification, and when choosing channel partners, the company does not wish to obtain orders through price competition or over-promising on delivery times.

The brain is smarter, can the hardware keep up?

The aforementioned person in charge of Yuanli Lingji mentioned that in the past, the industry focused more on whether software capabilities could keep up, but as models progress, hardware limitations have also become more obvious.

Robot local computing resources also need to handle tasks such as control and camera data processing; when deploying models, computing power, bandwidth, and heat generation need to be considered; more powerful chips may also significantly increase the total machine cost. He stated that the company is exploring more cost-advantaged chip solutions, and model inference costs will also affect actual deployment.

Cainiao's upgrade of the climbing robot this time also involves software and hardware; in addition to scheduling and path optimization, there are adjustments to the box-taking structure, equipment weight reduction, and battery performance improvements. Wang Zihao introduced that reducing the vehicle body weight can lower the load, striving for longer working time under the same battery conditions.

The climbing robot demonstrated by Cainiao

The hardware performance and usage cost of humanoid robots are also issues that have been repeatedly mentioned.

An industry insider, when communicating with reporters, cited dexterous hands as an example, stating that the durability and cost of dexterous hands will affect the deployment of humanoid robots; currently, the price of some dexterous hands accounts for about 1/3 of the robot, and the durability is also not entirely satisfactory, so in some tasks, enterprises are still adopting gripper solutions.

Ding Xia, using moving boxes as an example, talked about how a dedicated handling device might only cost a few tens of thousands of yuan and have higher efficiency, so why would customers use a more expensive humanoid robot to complete the same task? In her view, the grasping capability, stability, heat dissipation, and overall machine cost of humanoid robots still need improvement, and at the current stage, many applications still find it difficult to make the economic math work.

However, Ding Xia does not deny the future application space of humanoid robots. She believes that as technology matures, production increases, and prices drop, some scenarios will gradually meet the conditions for use. What she cares more about is, among the numbers announced by enterprises, how many devices truly enter commercial use, “commercialization is supposed to form repeat purchases.”Return to the previous page for more information