Amid the sweltering heat, the World Artificial Intelligence Conference (WAIC) unfolded in Shanghai earlier this month, showcasing around 25 different models of humanoid robots. Despite their flashy appearances, most were limited to static poses or simple interactive actions, with little to show in terms of practical applications.
Large models have catalyzed the popularity of embodied intelligence concepts. Humanoid robots provide a physical vessel for large models, bringing artificial intelligence into the real world. These models have enhanced humanoid robots’ perception, decision-making, and interaction capabilities, bolstering unsupervised learning and generalization frameworks.
However, remaining in the demo phase without practical applications means that even the most advanced humanoid robots cannot yet generate real-world value. While the technology has theoretically become more advanced, exploring real-world applications is ultimately more critical.
The pressing question now is how humanoid robots can transition from the lab to the market. Some potential applications have been mooted over the past six months:
- Agility Robotics’ Digit has made its way into Amazon’s warehouses, replacing human labor and able to operate continuously for up to 7.5 hours.
- Figure AI has inked an agreement with BMW to deploy the Figure 01 robot in its US manufacturing plant.
- Tesla’s Optimus is now part of its own workshops, reportedly engaged in battery sorting training.
- UBTech Robotics is also stepping up, collaborating with multiple carmakers to integrate its industrial humanoid robot, Walker S, into various automotive factory programs.
Deploying humanoid robots in automotive factories
Industrial manufacturing is a primary application area for humanoid robot developers. Compared to complex scenarios such as outdoors, homes, and supermarkets, the factory environment is relatively stable and simple, with ample data, providing fertile ground for early commercialization.
The development of humanoid robots can be divided into two stages. The first stage focuses primarily on showcasing the technology, with humanoid robots used in university research projects, school courses, and developer competitions.
Since last year, with the support of large models, the application of humanoid robots has moved toward the second stage, targeting real-world applications such as industrial manufacturing and logistics, and beginning to perform specific tasks. For example, in the US, humanoid robots from three companies—Tesla, Figure AI, and Apptronik—have begun training on the production lines of automotive factories.
Compared to the US, China has more companies developing humanoid robots. UBTech, Xiaomi, Leju Robotics, and Zhiyuan Robotics have all expressed intentions to apply humanoid robot technology in factories. However, only UBTech has announced the training progress of its humanoid robots in automaking so far.
Tesla: Two Optimus robots deployed in factories
In October 2022, when the humanoid robot Optimus made its debut, Tesla CEO Elon Musk proposed that future Optimus robots would work in Tesla’s factories, performing tasks like moving items and using wrenches to install bolts on cars.
Currently, the Optimus has been updated to the second generation, standing 1.7 meters tall, weighing 56 kilograms, and with a maximum walking speed of 5 miles per hour. It is equipped with actuators and sensor drives designed and manufactured by Tesla.
RELATED ARTICLE
In May this year, the second-generation Optimus started training in Tesla factories and can already perform tasks such as sorting batteries. According to official reports, Optimus, based on its own visual neural network and Full Self-Driving (FSD) chip, can sort electric vehicle battery cells and precisely place them into slots.
Currently, there are two Optimus robots deployed and operating autonomously, and Musk predicts that by 2025, there could be over 1,000 Optimus robots utilized across Tesla’s factories.
Figure AI: Targeting fully autonomous robots
Figure AI, a humanoid robot startup founded in 2022, has investors including OpenAI, Intel, Nvidia, Amazon, Microsoft, Samsung, and LG. Within a year, the company released its first humanoid robot, Figure 01. This robot stands 1.68 meters tall, weighs 60 kilograms, can carry a maximum load of 20 kilograms, has a maximum endurance of 5 hours, and a walking speed of 1.2 meters per second.
In January this year, Figure AI announced an agreement with BMW to introduce Figure 01 into BMW’s US manufacturing plant. According to the agreement, Figure AI will start by deploying a small number of humanoid robots and gradually cover more assembly line positions over 12–24 months.
For example, in a video released by Figure AI earlier this month, Figure 01 was given a positioning command to place the car baffle in the corresponding fixture and cover the two positioning pins that needed to be perforated. Based on GPT-4o, Figure 01 was not only able to understand the task and convert it into specific instruction steps to complete, but also adjust the direction of the car parts according to the angle of the fixture, correctly placing the parts in the appropriate positions.
According to official reports, Figure 01 uses an “AI-based visual model” that can achieve “fully autonomous control.”
Apptronik: Starting with delivery tasks
Apptronik, spun off from a lab at the University of Texas, was founded in 2016. The team partnered with NASA the following year to develop humanoid robots. In August 2023, Apptronik officially released its first humanoid robot, Apollo, which stands about 1.76 meters tall, weighs 72.5 kilograms, has a maximum load of 25 kilograms, and an endurance of 4 hours.
According to Apptronik’s CCO Barry Phillips, Apollo features unique design and technical characteristics. For example, its replaceable battery design allows it to continue working simply by replacing the battery, without long charging waits. Additionally, its force control architecture ensures safe and efficient task completion in warehouse and logistics scenarios.
In March this year, Apptronik announced a partnership with Mercedes-Benz to introduce Apollo into Mercedes-Benz’s factory in Austin, Texas. Apollo will participate in production line processes such as delivering parts to the production line for workers to assemble, inspecting components, and later delivering complete part kits.
UBTech Robotics: Walker S in production line training
Founded in 2012, UBTech is one of the earliest companies in China to develop humanoid robots and was listed on the main board of the Hong Kong Stock Exchange (HKEX) in December 2023. Its humanoid robot products focus on three main application scenarios: industrial manufacturing, commercial services, and home companionship. The Walker series is the most well-known among its products, having gone on display at Expo 2020 Dubai’s China Pavilion, the China Science and Technology Museum, and being exported to Neom, Saudi Arabia.
In 2023, UBTech’s humanoid robot Walker X entered factory training and testing, primarily used in SPS intelligent sorting and equipment aging testing. By linking with automated guided vehicles (AGV), Walker X was integrated into high-repetition, low-load production to enhance factory efficiency and process flexibility.
At the end of the same year, during UBTech’s listing ceremony on the HKEX, the industrial version of Walker S was unveiled for the first time, drawing attention as it struck the listing gong alongside founder Zhou Jian.
Walker S stands 1.7 meters tall, has 41 high-performance servo joints, and is equipped with multidimensional force perception, multieye stereo vision, omnidirectional hearing, and inertial, ranging, and other perception systems. Its hand features interchangeable end effectors, enhancing its efficiency in various factory scenarios.
In February this year, in a video released by UBTech, Walker S was shown performing tasks on the production line at the final assembly workshop in Nio’s Hefei base. The robot was shown undertaking various tasks, including door lock inspection, headlight cover, board inspection, seat belt inspection, and car badge application. These tasks covered multiple production stages, and Walker S could also cooperate with human employees to complete car assembly and quality inspection work.
Subsequently, UBTech reached cooperation agreements with Dongfeng Liuzhou Motor and FAW-Volkswagen (Qingdao). By establishing more production lines and factories, they aim to enhance the tool operation and task execution capabilities of humanoid robots, further accelerating their application in manufacturing scenarios.
Zhou said that Walker S will undergo extensive factory testing in 2024 to prepare for large-scale deployment of humanoid robots in the next two years.
As of mid-2024, the implementation of humanoid robots in industrial manufacturing is still in the training phase. However, industry practitioners are optimistic, predicting that preliminary implementation could be achieved by the end of this year.
Moving from lab to market
The innovation of a disruptive technology is a gradual process. Compared to other scenarios, the industrial manufacturing environment is relatively simple and controllable. It is usually driven by efficiency and precise cost control. Factory workshops are closed, structured environments that can provide a more stable working environment for humanoid robot applications.
On the other hand, after years of development, the industrial automation industry has matured. Equipment like heavy-duty robotic arms and collaborative robots have accumulated a large amount of real-world scenario data, which can be directly reused in the training of humanoid robots. This data allows humanoid robots to continuously iterate and adapt to scenario requirements.
The value of industrial manufacturing scenarios is continuously rising, and almost all humanoid robot developers mention it when discussing application implementation. Automakers also hope to integrate the best technologies, introducing humanoid robots into production. However, for humanoid robots to truly perform in car factories, there are still many issues to be resolved.
Jiao Jichao, executive director of UBTech’s research institute and head of the humanoid robot team, said that his main task this year is to communicate with automotive manufacturers about the application of humanoid robots. He has currently researched the needs of more than a dozen domestic car manufacturers.
He discovered that, in automotive manufacturing scenarios, there is no need for humanoid robots to demonstrate single movement abilities such as running or jumping. Instead, there is an emphasis on practicality, such as the intelligence, dexterity, stability, scalability, and cost-effectiveness of the robots.
In the past, the technical path of traditional robots revolved around single scenarios with basic algorithm capabilities. Through pre-programming, they could recognize simple environments and complete simple tasks.
However, in real-world scenarios, humanoid robots need to have autonomous decision-making abilities to deal with unexpected events. This requires a high level of intelligence and the ability to understand and handle multiple tasks. The development of computational power and the maturity of large models have played an important role in this.
For example, Tesla’s Optimus uses FSD technology, which has been maturely applied in EVs. When applied to Optimus, it uses FSD’s end-to-end large model for high-speed inference. After cleaning enough user data and feeding it to the model, it can enable faster response times and more agile decision-making.
OpenAI built an embodied intelligence model for Figure 01 based on GPT-4o. This model endows it with the ability to perceive, recognize, and understand objects and natural language from a brain level, enabling it to perform logical reasoning and autonomous decision-making based on task instructions.
UBTech adopts an architecture that combines a large model and a small model. The large model acts as the “brain,” processing multimodal information, understanding scenarios and tasks, and providing planning. The small model executes specific tasks. This solution leverages the understanding and generalization advantages of large models and the low energy consumption and task-specific effectiveness of small models, meeting various task requirements of humanoid robots.
From the hardware perspective, humanoid robot companies have invested significantly in developing core components like servo drives. For example, UBTech has developed a harmonic servo drive specially for industrial scenarios. It features high torque density, small size, lightweight, and minimal backlash, supporting high-precision dexterous operations. This allows humanoid robots to walk stably in complex and changing environments while enabling their arms to handle heavy tasks.
Tesla developed a linear drive for Optimus, integrating servo motors, reducers, lead screws, sensors, and integrated motion units. This achieves precise speed, position, and force control, granting Optimus more flexibility during operation.
Regarding upper limbs, all humanoid robots currently involved in factory training have made certain progress in dexterous hand operations. For instance:
- The second-generation Optimus is fitted with fingers that have tactile sensing and 11 degrees of freedom, capable of handling fragile items like eggs. In May, Musk revealed that, by the end of this year, Optimus’ dexterous hand freedom will increase to 22 degrees, with actuators placed in the forearm to simulate human work methods.
- Figure 01’s hand features four fingers and a foldable thumb, with six degrees of freedom, allowing it to open doors and use tools.
- Apollo’s hand can use single-degree-of-freedom end effectors or different combinations of effectors in addition to a five-finger hand to complete tasks.
- UBTech’s Walker S uses self-developed arms that can perform flexible operations and complete heavy-load tasks by changing different end effectors. Its servo drive designed for industrial scenarios enables high-precision position control, allowing Walker S to engage in fine tasks like sorting, inserting, and screwing.
Integrating humanoid robots into long-term operations stably demands high levels of technological accumulation from companies. Additionally, automakers have strict requirements for data security and governance. For companies that do not fully develop their technology in-house, outsourcing to third parties can significantly increase the risk of data leaks.
Taking UBTech as an example, as an early entrant in China, it has independently completed the full-stack development of its humanoid robot technology. Addressing the actual needs of industrial scenarios, UBTech has not only mapped out algorithms but also developed core components in-house, continuously iterating technology and optimizing processes to solve real-time feedback issues, ensuring operational stability in real-world scenarios.
From the training work of Walker S in factories, its assembly and quality inspection tasks showcase the integration of perception, hand-eye coordination, positioning and navigation, and gait control technologies with hardware and scenario refinement.
Even smaller companies like Apptronik, with participation from institutions like NASA, have accelerated breakthroughs in bipedal motion, perception, and grasping capabilities, meeting Mercedes-Benz’s diverse task requirements.
Mastering specific scenario know-how is also a critical factor affecting the performance of humanoid robots. Currently, humanoid robots are being explored in a highly targeted fashion, requiring companies to have a thorough understanding of a single scenario by collecting sufficient data to accurately grasp the actual needs of the industry.
In this regard, Tesla has a natural advantage. Its automotive factory has become a mature application testing ground for Optimus. Optimus undergoes daily training in a real factory environment, collecting data and continuously improving performance iteratively, closely resembling the process of training human employees.
UBTech formed a joint venture with Miracle Automation, an automaking supply chain company, providing a critical entry point into the field. This makes UBTech one of the few humanoid robot companies that does not produce cars but participates in car manufacturing. The range and depth of its partnerships with car companies are broader and deeper. Currently, UBTech has announced cooperation with three automotive companies, obtaining real-world scenario data for humanoid robots.
Finally, cost control capability is an important factor in commercializing humanoid robots. Ultimately, new technology needs to offer meaningful benefits to encourage user and business adoption.
Localization of the supply chain helps humanoid robot companies achieve cost optimization. Hardware suppliers in China have robust mass production capabilities. The controllers, sensors, and battery systems used in humanoid robots overlap significantly with those in industrial robots and new energy vehicles. By strategically repurposing some pre-existing supply chain resources, the overall cost of humanoid robots can be effectively reduced.
Jiao said that UBTech has increased the localization rate of its core servo drive components from 40% a few years ago to 90–95% today. Domesticization rates for high-power and low-power servo drive components have respectively reached 90% and 95%, significantly reducing the cost of humanoid robots. Currently, the cost of UBTech’s large humanoid robots has dropped to USD 40,000–50,000.
Tesla has not yet announced the cost of Optimus but continues to select supply chains in China. Musk has said before that the aim is to keep the prices of its humanoid robots at around USD 20,000 in the future. Based on industry data, considering cost control requirements, Chinese manufacturers are expected to leverage cost advantages to enter Tesla’s supply chain.
Another favorable condition for Chinese manufacturers is that governments in Beijing, Shanghai, Shenzhen, Hangzhou, and other cities are actively promoting collaborations across the humanoid robot industry chain.
For instance, in Beijing, UBTech, Xiaomi, Jingcheng Machinery Electric, and other leading companies established the Beijing Humanoid Robot Innovation Center last year. This center is dedicated to the R&D of key technologies in the industry, creating platforms to promote overall industry development.
Currently, in the Beijing Economic-Technological Development Area, there is not only a concentration of humanoid robot companies like UBTech and Xiaomi but also key component producers like Jingcheng Machinery Electric, Tsino Dynatron, RobStride, and Pixelcore, forming a collaborative ecosystem.
Additionally, automakers consider the qualifications and experience of suppliers when selecting partners, emphasizing whether humanoid robot companies can provide long-term and reliable services. This is because the usage and maintenance cycle of production equipment purchased by automotive factories is usually 3–10 years. Therefore, in choosing partners, whether companies can be relied upon to provide long-term after-sales services is crucial, posing a high threshold for most current humanoid robot startups.
Humanoid robots must ultimately move from the laboratory to real-world scenarios to realize their true value. Industrial scenarios, represented by automotive manufacturing, have presented a pathway for the commercialization of humanoid robots. Although humanoid robots have not officially been integrated, observing the companies that have started applications in this area provides a glimpse into the necessary qualities for commercial application of humanoid robots.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Xiao Xi for 36Kr.