on the eve of the humanoid robot blowout | ubtech qin wenlong: large-scale production of humanoid robots will be in the next 2 to 3 years
2024-08-29
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
on the eve of the humanoid robot boomlast october, the ministry of industry and information technology (miit) issued the “guiding opinions on the innovation and development of humanoid robots”, which gave high expectations for the development of humanoid robots, saying that they are “expected to become a disruptive product after computers, smart phones, and new energy vehicles, which will profoundly change human production and lifestyle and reshape the global industrial development pattern.”humanoid robots are experiencing a transition from landing to mass production. red star capital interviewed a number of manufacturers and discussed with them the eve of the blowout of china's humanoid robot industry.
with the development of humanoid robots, it is becoming a reality for them to work in factories. ubtech's walker s has entered the smart automobile factories of byd (002594.sz), nio (09866.hk; nio.us), and dongfeng liuzhou motor, collaborating with humans on the assembly line to complete tasks such as assembly and quality inspection.
"humanoid robots are in a rapid development stage, and it is very fast to go from 0 to 1, but the step from laboratory models to factory applications must be taken well. what factories need are mature and stable products, not laboratory-level prototypes," said jiao jichao, vice president of ubtech robotics and executive director of the research institute, in an interview with the media.
qin wenlong, senior engineer of motion control algorithm at ubtech robotics, said in an exclusive interview with red star capital bureau that data accumulation is an important consensus in the development of humanoid robots. he believes that the time for humanoid robots to leap from small-scale mass production to large-scale production may occur in the next two to three years.
humanoid robots focus more on industrial scenarios
a lot of data still needs to be accumulated
red star capital bureau: how is ubtech's humanoid robot application development progressing?
qin wenlong:now is the poc (proof of concept) stage, all robots enter the real car factory, complete the practical training tasks, and make preliminary preparations for large-scale implementation in the future. at this stage, we will discuss some scenarios with customers, such as handling, quality inspection, sorting, etc. after determining the scenarios, we will do skill development and joint debugging internally, and enter the car factory to complete the poc after passing the test. after one round is over, we will evaluate the implementation effect and then determine the implementation direction of the next round of poc. the ultimate goal of our poc is to accumulate practical training skills and prepare for mass delivery to the factory.
we are now focusing more on industry, which is a more structured scenario that is more suitable for humanoid robots to be deployed at this stage. the scope of work and the content of tasks are relatively certain. the environment of home scenarios is relatively complex, and whether it is housework or interaction with family members, the requirements for robots are very high.
red star capital bureau: what is the consensus on technology? what variables does the big model bring to humanoid robots?
qin wenlong:humanoid robots now have a relatively clear technical route, which is to accumulate a large amount of data. embodied intelligence emphasizes universality. to achieve universality, a large amount of data is needed, and humanoid robots must learn skills on their own.
data is important for artificial intelligence, and the same is true for humanoid robots. we have applied deep learning, imitation learning, and reinforcement learning methods to humanoid robots, and have shown the effect of skill generalization, such as the ability to resist interference and the ability to disassemble tasks. however, the current robot skill learning is still in its early stages, because humans are still very complicated when doing this, so we still need to continue to accumulate a large amount of data in the future, so that the scaling-law of artificial intelligence can also be mapped to humanoid robots to achieve a significant improvement in the level of intelligence.
the changes brought by the big model are obvious interactive capabilities and task decomposition. before the big model, the robot needed to be pre-programmed for each step. now the big model plays the role of a good brain, decomposing tasks and interacting with people. the development of the intelligence level of humanoid robots will advance in sync with the development of artificial intelligence and big model technologies. these two technologies are strongly related.
red star capital bureau: are the technologies and experiences applicable across all scenarios?
qin wenlong:the model changes between different scenarios are not that big, but more about data. for example, in industrial scenarios, only relevant data of industrial scenarios are collected, and the trained model is a worker; if a domestic service worker is needed, the data "fed" is related to that scenario. data is more critical than models. the model itself has a certain degree of versatility. through the same model, it can be trained to be a worker who can screw screws, and it can also be trained to be a nanny who can do housework.
the data collection process itself is not difficult, but it takes time to accumulate data. first-line data is definitely the best, and we will also generate and accumulate data in simple scenarios and simulation scenarios. but it is very important that you must know clearly what scenarios the humanoid robot will be used in, so that you can know what data to collect now.
the development of humanoid robots at home and abroad is "moving forward in parallel"
domestic enterprises have the advantage of application scenarios
red star capital bureau: there are many robots in industrial scenarios. what can humanoid robots do?
qin wenlong:the advantage of humanoid robots over industrial machines and collaborative arms is their versatility. now we only use them to screw, carry, or perform quality inspection, but in essence, one machine can achieve three functions.
when we went to investigate companies, we found that there are a large number of welding robots and agvs (automated guided vehicles) on automobile production lines, but there are still a large number of workstations that need to be done manually, especially some jobs that even require standing or waiting for a long time. currently, industrial robots cannot automate these tasks very well, which may involve certain modifications to the site. for example, to realize the automation of the entire logistics, the entire workshop needs to be modified, which is a huge cost investment for the company and also involves great risks.
to meet the requirements of universality and to be able to be deployed quickly without extensive changes to the production line, humanoid robots have advantages in such workstations. humanoid robots can be piloted on a small scale to quickly see the effect.
humanoid robots are not competing with existing industrial robots and collaborative robots, which have their own scope of application. for humanoid robots, it is in the early stage of industrial implementation, and it is not yet at the stage of "volume", and everyone is still trying.
red star capital bureau: we see that many domestic companies will benchmark tesla's optimus prime in international competitionubtechwhat are the advantages?
qin wenlong:overseas countries are also in the initial stage of technology verification. on the application side, we have the advantage of application scenarios in china, and we have more manufacturing companies. on the technical level, i think it is more of "moving forward together" or each has its own advantages, and there is no technological generation gap. tesla optimus only put the robot in the car factory and performed a single action such as sorting.
the implementation of humanoid robots is a very complicated matter. every company is doing it in stages, starting with small-scale and small batches of dozens or hundreds of units, and then to large-scale production of tens of thousands of units.
for us, we will not train all skills. instead, we will use poc to understand customers, scenarios and needs, and train accordingly based on what skills are needed. in this way, we can advance to the next mass production stage. in my opinion, the time span between the two stages may be only 2 to 3 years.
foreign countries and we are also competing with each other in this regard, and there is still a certain gap in completing high-level multi-tasking and meeting the complete needs of customers.
at this stage, everyone has not yet fully focused on the application scenarios. as the party closest to the application scenarios, domestic enterprises need to seize our own advantages and focus on finding scenarios.
red star news reporter wang tian
editor: deng lingyao
(download red star news and get a reward for reporting!)
