Robots can perform tasks now, but the industry still needs more cunning "villains"
"Can you keep the cycle time within 15 seconds? What is the accuracy range? How long is the battery life? Can it automatically swap power supplies? Can it handle a 10-kilogram payload?" On July 18, a business leader in charge of the CNC (Computer Numerical Control machining) operations at a manufacturing plant in Anhui Province asked several questions after watching a robot demonstrate box-moving tasks. The staff member conducting the demonstration could not answer and turned to the nearby technical team for assistance.
Similar conversations frequently took place at the 2026 WAIC (World Artificial Intelligence Conference) exhibition hall. This year's attendees are no longer satisfied with watching robots sing and dance; they are more concerned about whether robots can perform actual work and how capable they are at doing it.
In Pavilion H3 on the second floor of WAIC, more than half of the 160 embodied intelligence companies are trying to prove one thing: robots can truly perform practical work, showcasing capabilities such as making tea, folding clothes, moving boxes, and organizing items. Several embodied intelligence companies valued at tens of billions of yuan have brought entire scenarios from automobile factories, logistics sorting facilities, 3C manufacturing lines, and smart pharmacies into the exhibition hall.
01
Scenarios Expanded to More Than a Dozen
Galaxy General is a leading humanoid robotics company. In previous exhibition demonstrations, Galaxy General was keen to show robots performing full workflows such as fetching drinks and retrieving medications for humans. This time, Galaxy General demonstrated its robots' capabilities in screwing on new energy battery outer casings and moving 30-kilogram boxes.
A Galaxy General staff member stated that this batch of heavy-duty robots has been working at CATL's factory for three months, marking the first large-scale deep deployment of humanoid robots in the new energy manufacturing sector.
Cross-Dimensional Intelligence demonstrated robots playing chess and performing mobile phone assembly tasks, with the robots operating fully autonomously without remote control. Noticing that Lingchu Intelligence across the aisle was also showcasing robots grabbing mobile phones and placing them into packaging boxes, the two companies had identical ideas. Cross-Dimensional Intelligence directly moved its mobile phone assembly workbench directly opposite to its competitor, hoping for a "head-to-head showdown."
The humanoid robot on display at Stone Intelligent Navigation's booth worked on a panel covered with dozens of thin wires. It needed to pick out one wire from the cluster, organize it, and insert it into a tiny hole. This scenario is primarily found in automotive wiring harness factories and has already been deployed in real-world settings.
Ding Wenchao, Co-Founder and Chief Scientist of Stone Intelligent Navigation, stated that the speed and accuracy of robots working in automotive wiring harness factories are approaching human levels. In a real factory, humans handling dozens of wires at once is a tedious task, but robots do not experience fatigue. Furthermore, robots can perform tasks that humans cannot, such as inserting wires simultaneously with both hands, which could potentially double efficiency compared to humans.
ZhiYuan, which live-streamed its robots working at the Longcheer factory in June this year, brought the entire production line where its robots operated at Joyson Automotive's auto parts factory to the exhibition booth. ZhiYuan demonstrated the collaborative capabilities of three robots working together at Joyson's factory: one robot handles loading and unloading, one is responsible for packing boxes, and one moves the boxes. A ZhiYuan staff member explained that this demonstration was primarily designed to showcase inter-robot collaboration.
Lingxin Qiaoshou, a company specializing in dexterous hands, has a few square meters of area at its booth where a dexterous hand picks up several mechanical fingers from a table and assembles them into a complete robotic hand.
Zhang Yanbai, Co-Founder of Lingxin Qiaoshou, said that compared to last year, the flexibility and service life of dexterous hands have been significantly improved. In terms of flexibility, last year's dexterous hands could only press down on piano keys with concentrated finger force, but this year's five fingers can now flexibly pluck piano strings. Regarding lifespan, last year's best domestic dexterous hand achieved a maximum of 500,000 opening and closing cycles, while Lingxin Qiaoshou can now reach 2 million cycles.
According to incomplete statistics, the 160 embodied intelligence companies at WAIC collectively showcased more than a dozen scenarios, including clothes folding, laundry, box moving, box transportation, express sorting, loading and unloading, coffee making, breakfast preparation, bottle cap opening, item storage, drink vending, tea pouring, energy facility inspection, and metrology testing.
Deploying robots into households is a more challenging scenario than factory deployment, and very few companies dared to attempt the home sector in previous years. However, this year, multiple embodied intelligence companies showcased robots designed for household scenarios at WAIC.
The unified characteristic of these robots is their small size and cute, endearing appearance. The Qiyuan Q1 launched by SWEET is a representative product, standing 88 centimeters tall—less than the height of an adult's thigh—weighing about 15 kilograms, and priced at tens of thousands of yuan. The product focuses on companionship features, such as speaking English with children and playing with them. Tian Hua, CEO of SWEET, stated that its target user groups are developer geeks, tech trendsetters, and families with parent-child companionship needs.
02
Driving Upstream Industries to Profit
As of July 2026, the number of robots that have entered factory operations is not large. Reporters learned that individual companies have deployed at most a few hundred units, while some have only a few, with most still in small-scale pilot phases. However, the demand from embodied intelligence companies for robots with practical work capabilities has already driven a group of upstream suppliers to generate profits.
Tashan Technology is a developer of artificial intelligence tactile sensing chips and application solutions. Fu Yihui, Vice President of Marketing and Ecosystem at Tashan Technology, told reporters that at the beginning of last year, less than 20% of the dexterous hand industry required tactile sensors, but this year that proportion has risen to 60%. Without tactile sensors, training robots to perform work tasks would be costly and ineffective. In the first half of this year, Tashan Technology's order volume quadrupled compared to the entire previous year, delivering tens of thousands of tactile sensors monthly, with projected year-end orders expected to be over 10 times last year's total.
Reporters learned that during WAIC, another tactile sensor company, Yimu Technology, secured 1 billion yuan in financing, reaching a valuation of over 10 billion yuan. Its clients are primarily data collection companies and robotics enterprises.
Data companies have also seen revenue growth. In the first quarter of this year, physical AI data firm Guanglun Intelligence received 550 million yuan in orders, and the company currently has a valuation of 15 billion yuan.
Yang Haibo, Co-Founder and President of Guanglun Intelligence, told reporters that this year's data demand from the embodied intelligence industry is a thousand times greater than last year. Last year, embodied intelligence companies only purchased hundreds or thousands of hours of data, but this year's demand has surged to millions of hours. He believes that humanoid robot companies are rushing to deploy in real scenarios this year but are facing issues of data scarcity and insufficient deployment capabilities, leading them to purchase large volumes of data.
Two additional changes have emerged in the embodied intelligence data sector this year: First, the repurchase rate among embodied intelligence companies has increased. Previously, most enterprises purchased data through one-time contracts, but this year, repeat purchases have become common. Second, the repeat sales rate for data firms has risen, allowing them to sell the same dataset to different clients.
03
Bottlenecks Still Persist
The questions raised by the Anhui factory business leader did not ultimately receive satisfactory answers. She told the Economic Observer that her factory wants robots to work in high-temperature, high-risk environments to improve manufacturing efficiency. At the same time, they are unwilling to spend hundreds of thousands of yuan on a robot whose work efficiency is less stable than that of humans. The robot exhibitors she visited that day could not yet demonstrate capabilities that match her requirements.
While many robots at the WAIC exhibition showcased work capabilities across numerous scenarios, their operating speed and efficiency are generally not high. One attendee filmed a video of a robot folding clothes for two minutes, but the robot still had not finished. Disappointed, he put down his phone and said, "I'd rather fold it myself."
Furthermore, significant hardware bottlenecks remain for robots. Whether it is dexterous hands or the robot's main body, their limited service life makes them uneconomical for factory use.
"Compared to seasoned factory workers, robots' work success rate and efficiency are certainly not as good. But the clear change this year is that robots have truly entered factories," Zhang Yanbai, Co-Founder of Lingxin Qiaoshou, remarked. He believes this means robots can now access large volumes of factory data and learn to perform tasks like apprentices, which is a promising beginning.
Yang Haibo predicts that as embodied intelligence startups increase investments in robot hardware, and manufacturing-savvy companies such as BYD and Honor join the sector, robot hardware prices will drop significantly next year while performance improves. By then, factories purchasing robots for work will be an investment that clearly reduces costs and increases efficiency. "This year, everyone's expectation for robots is to validate factory scenarios; calculating profitability is a task for next year."
04
More "Tricky Troublemakers" Are Needed
Exhibition booth staff are on tenterhooks, fearing that any robot malfunction spread through social media could turn into a public relations crisis. Attendees propose numerous demands that gradually expose robots to the complexities of the physical world. For example, while a ZhiYuan robot is tightening screws, attendees may ask to reposition the screws at different angles or move them mid-operation; when playing chess, robots must account for humans retracting their moves.
On July 19, Lu Cewu, Co-Founder of Qiongche Intelligence, stumped a robot by placing multiple plastic water bottles side by side tightly together as the robot attempted to grasp them. Due to the thickness of the mechanical gripper, the robot could not find a gap to grab between the bottles, and attempting to grasp from the other side exceeded its operational range. He suggested that robots should learn to actively rearrange and scatter the bottles before attempting to grasp them.
A representative of the robot company that was "thrown a curveball" said that humans going out of their way to "make things difficult" for a robot is for its own good, with the goal of improving the robot's generalization ability and robustness.
A representative from Cross-Dimensional Intelligence told the Economic Observer that when training robots, R&D personnel must introduce more failure cases to help them learn recovery mechanisms from errors. The richer the training data, the more "textbooks" the robot has "memorized," and the more scenarios it can handle.
On one hand, when developing foundational generative universal models such as pre-trained models, teams use a wide variety of data sources, including human videos, internet videos, and Ego (first-person perspective) data without corresponding physical hardware. On the other hand, interference is introduced during post-training for specific commercial scenarios, with teams implementing targeted optimizations whenever issues such as model clipping or objects being knocked away occur.
Scenarios involving direct human interaction are the most challenging, because humans are the "troublemaking variable" that introduces significant uncertainty. Logistics sorting, coffee-making robots, and similar scenarios have relatively fewer environmental variables or human interference, making them more suitable for commercial deployment.
An investor in an embodied intelligence company believes that the current lack of sufficient intelligence in humanoid robots is a major bottleneck plaguing the embodied intelligence industry. The explosion in AI product capabilities originated with the launch of ChatGPT in 2022. However, the embodied intelligence industry has not yet experienced the emergence of intelligent capabilities, and practitioners are still waiting for the "ChatGPT moment" of the embodied intelligence sector.
Robots carry public expectations of entering households and providing elderly care services. Behind the scenes of WAIC's robot work scenario demonstrations, most exhibitors began collecting on-site data and fine-tuning their programs as early as July 14 in preparation for the conference. A senior executive from a leading embodied intelligence company explained that environmental interference is fatal to robots—for example, changes in on-site lighting or attendees using phone flashlights could cause the visual model to crash. "Deploying directly using a world model is still too difficult; no company has mastered it yet. The real-world deployment of robots must rely on engineering-level anti-interference testing."
This article is from the WeChat Official Account "Economic Observer", written by Ren Xiaoning and Chen Yueqin, and published with authorization from 36Kr.