The Present and Future of Robots: "Embodied Intelligence" may push the boundaries between man and machine

21st Century Business Herald reporter Lin Dianchi reported from Shenzhen

On December 19, the drone landed slowly at the Shenzhen International Graduate School of Tsinghua University and delivered a cup of freshly squeezed juice, which is one of the cases of robots entering public life. It is understood that this is also Meituan's first college route, and the take-off point is set in the Yitian Holiday Li business district, 2.6 kilometers away, and it generally only takes about 6 minutes to deliver the meals after they are packaged.

In 2023, robots are moving from the original industry to the service scene, people have become accustomed to hotel food delivery robots, sweeping robots have entered thousands of households, and drone delivery has taken the lead in landing in some cities.

According to data provided by Meituan, as of November 2023, 22 routes have been opened in 8 business districts in Shenzhen and other cities, covering various scenarios such as community office buildings, scenic spots, and hospitals, and a total of more than 210,000 user orders have been completed.

On the other hand, with the advancement of AI technology and the emergence of large language models such as ChatGPT, human-machine communication is more fluid and seamless than ever before. In addition, if the robot sensor is given, with the blessing of artificial intelligence, the robot may be able to perceive and interact with Xi like humans, which is the hot "embodied intelligence" at the moment.

Now: Opportunities and challenges

Recently, Tesla CEO Elon Musk released the latest video of the humanoid robot Optimus, which revealed that the second-generation Optimus robot (Gen 2) has a 30% increase in walking speed, improved balance and body control, and because the Optimus ten fingers are equipped with finger tactile sensors, the robot can pinch eggs with two fingers.

At the robot roundtable forum held at the same time on the same day, the participating experts all believed that the progress of humanoid robots this year was eye-catching, and products with high completion of humanoid robots have emerged at home and abroad.

In terms of landing applications, "this year, collaborative robots in the industrial field and AGV robots in the logistics field have made considerable progress. Zhang Jianwei, a foreign academician of the Chinese Academy of Engineering, said.

In addition, statistics from Yiou think tank show that in 2023, the output of service robots will increase by 47.6% year-on-year, ending a year-on-year decline for 12 consecutive months.

Major breakthroughs have been made in the R&D and production of core components such as vision sensors, SLAM systems, AI robot chips, and robot operating systems. The localization of these core components not only reduces the manufacturing cost of service robots, but also helps to promote the competitiveness of Chinese service robots in the global market.

For example, Pudu Technology has achieved industry-leading technological breakthroughs in key areas such as robot operating systems and SLAM systems, which have greatly improved the deployment and autonomous navigation capabilities of service robots.

It is understood that in the past year, Meituan has made progress in the exploration of robots such as automatic delivery vehicles. For example, since the implementation of the automatic delivery vehicle delivery service in March 2018, it has completed technical testing and trial operation in many places in China such as Xiong'an New Area, and has served more than 100 communities in Beijing, Shenzhen and other cities, with autonomous driving mileage accounting for more than 99%.

However, service robots still face challenges in commercial applications. Mao Nian, head of Meituan's drone business and chairman of Meituan Robotics Research Institute, said that the first thing is that the purchase of robots requires high initial investment costs and the actual utilization rate of robots is not high. In addition, the time cost of operation and maintenance costs and maintenance that the robot cannot be used normally is relatively high, and the robot is often only good at doing one thing, and the application of a single robot in multiple scenarios is relatively rare, and if there is a problem with any part or any chain, the user experience and perception will decline.

Challenges are often accompanied by opportunities, and the key is whether the market demand can be effectively tapped. Zhang Hong, an academician of the Canadian Academy of Engineering, said that the development of robots should be judged from two dimensions: business model and technological progress. For example, DJI has been recognized by the market for its aerial photography and has succeeded in the field of drones, while Meituan has already occupied a place in food delivery, and the promotion of drone delivery through the platform is also in line with the market demand.

Future: Embodied intelligence breaks through the boundaries between man and machine

Although large language models are expected to revolutionize the field of robotics, they still fail to fully meet the public's expectations of robots in terms of comprehension, association, and interaction.

In 1950, Alan Turing, the "father of artificial intelligence", first proposed the concept of "embodied intelligence". The basic assumption is that intelligent behavior can be Xi by agents with corresponding forms by adapting to the environment.

"Embodied intelligence is not only about bipedal robots, but also about the ability to understand the world, and form is just a carrier. Xu Huazhe, head of the Embodied Intelligence Laboratory of Tsinghua University, said.

With the development of robots, the industry generally believes that embodied intelligence will be the future direction of robot development. However, how to make up for the gap between them and human intelligence, robots still have gaps that cannot be ignored in terms of logic, perception, and human-computer interaction.

The exponential growth of artificial intelligence has brought strong impetus to the development of robots, and the next step is to take multimodal embodied intelligence as the main body of practice for future development.

Zhang Hong said that to put it mildly, robots have not undergone qualitative changes in the past 40 years, and one of the keys to solving the problem of robots is how to instill existing knowledge into robots, and large models will be a good entry point.

Early research on embodied intelligence found that it was incredibly difficult to generalize even agents to tasks that closely resemble training data.

For example, the strategy of "pushing an object from the middle of the table to the upper right corner" command does not work for the task of "pushing an object from the middle of the table to the bottom right corner" because the upper right and lower right corners of the table are not the same at the image recognition level, and therefore are not the same in the "eyes" of the robot.

After the pre-training of the embodied intelligent large model, it is expected to achieve a significant improvement compared with the direct training.

Xu Huazhe said that generative simulation of embodied intelligence is still relatively expected, and using this training method, the model has learned such an abstract concept as "putting A into B", which is expected to form the ability to systematically generalize to new tasks.

Experts at the meeting believe that in the next three years, embodied intelligence will make breakthroughs. Embodied intelligence will be the first to land in service scenarios, followed by industrial scenarios that require high frequency and are highly dependent on engineers.

For more information, please download the 21 Finance APP