Many experts advise to improve the interaction of service robots
Infant service robots urgently need to master correct learning methods
In the equipment exhibition area of the 2nd CIIE, a 5-finger manipulator service robot is performing an action demonstration. Photo by Xinhua News Agency reporter Fang Zhe
"Service robots are a carrier of artificial intelligence. They have different functions for different scenarios, such as industrial scenarios, elderly care scenarios, and special scenarios." At the 22nd China Association for Science and Technology Annual Conference held recently, Robots and Artificial Intelligence At the Industry Development Forum, Professor Han Zitian of Macau University of Science and Technology and Chairman of Macau Society of Artificial Intelligence and Robotics said.
"When I was studying, the price of a robot was about 600,000 to 800,000 yuan. 10 years ago, 300,000 yuan has now become 150,000 yuan. How much is the domestic price? About 80,000 yuan ." Wang Tianmiao, honorary director of the Institute of Robotics of Beijing University of Aeronautics and Astronautics, said that the popularity of robots has become faster and faster.
According to Yiou's "2020 China Service Robot Industry Development Research Report", breakthroughs in artificial intelligence technology and the reduction of core component costs have accelerated the penetration of service robots in various fields. In the past five years, the growth rate of China's service robot industry has been higher than the global average growth rate, and the market scale accounts for more than 25% of the global market. At the same time, it has global competitive advantages in the industrial chain and industrial environment. On the basis of sustained rapid growth, China's service robot industry will continue to expand rapidly in the future with huge potential.
However, experts generally said that the current service robots are equivalent to infants and young children in terms of intelligence, especially autonomous interaction. This is not only an obstacle to the industrialization of robots, but also a place where the main force should be exerted in the future.
Poor interactivity limits the development of service robots
Wang Tianmiao cites several directions in which artificial intelligence robots may have broad application prospects in the future: such as robotic agents in the field of enterprise services; elderly care, surgical robots, intelligent prostheses, etc. in the medical and health field; security, fire protection, and environmental protection in the field of smart cities , And logistics, driverless, financial technology, etc. "A large number of robots using artificial intelligence will support efficient and abundant material production activities and the reconstruction of production factors. In this context, let's discuss the future development pattern of artificial intelligence and robots." Wang Tianmiao said.
"The biggest advantage of robots is that they can do some repetitive, dangerous or disgusting tasks for humans, such as security work. If people are unwilling to patrol or be on duty at night, these positions can be replaced by service robots. So it cannot be said that service robots are completely To compete with humans, service robots are more to reinforce and enhance human work." Han Zitian said.
However, in Wang Tianmiao’s view, the current environment suitable for the work of service robots is still relatively fixed, the operation process is standardized, and the work is heavy, monotonous and boring; and in terms of early childhood education and nursing, and nursing for the elderly, service robots are still "unhelpful. ". According to Jens Kober, an associate professor at Delft University of Technology in the Netherlands, what robots can do with artificial intelligence is mainly related to cognition and advanced reasoning, but robots currently do not have these capabilities. In addition, Further development is needed in terms of actual movement and interaction with the environment.
Han Zitian also said that robots currently have no thoughts or autonomous consciousness. Although it can perform visual perception, voice perception, etc., in the final analysis, current robots do not have autonomous thoughts, nor can they make good autonomous decisions, have no thought and knowledge system, and cannot interact well with humans.
Infant service robots rely more on scenes
"Service robots are like newborn babies. Only when many new technologies such as vision technology and language technology are mature, will there be better service robots." Han Zitian said, and general-purpose service robots are currently difficult Achieved.
"So, the development of service robots so far depends more on scenarios. This is very similar to industrial robots. Industrial robots can only be used when the production process is clearly defined. The same is true for service robots, and their service scenarios must be clearly defined. , Make a work plan, the service robot can be used in the work scene." Han Zitian said. He emphasized that this is the scene dependence of service robots. Only by clearly defining the scene and the behavior of the robot can people make a more useful service robot.
Wang Yinxue, vice president of Megvii Technology, gave an example of an AI stacker. In a three-dimensional warehouse, no matter how tidy the goods are, they will inevitably fall, which often causes the goods to be crushed. After installing a camera on the stacker, not only can you observe the falling of the goods, the AI system can also calculate where it fell from through the trajectory, and predict the falling situation. In addition, the stacker is not only equipped with a camera, but also equipped with far-infrared detection equipment, patrolling at any time, if the local temperature rises above a certain level, it will alarm to prevent fire. This kind of fixed scene, fixed mode artificial intelligence machine application, the effect is immediate.
"The important difference between robots and smart devices is that robots have behaviors, and we even hope that they will have autonomous behaviors in the future." Han Zitian said that what we are talking about depends on scenes, such as elderly care scenes, car scenes, medical scenes, and security scenes. It is equivalent to the boundary of the robot service. What is important is its behavior in this service boundary.
Reinforcement learning can improve robot intelligence
“In many designs in the past, robots were very passive. They used voice to inquire about the robot’s situation, or asked questions about specific scenarios with the robot, and the robot’s answering process was standardized. Nowadays, robots have more and more social functions. Scenarios such as serving nursing homes, or interacting with children, active interaction has become a new hot spot.” According to Jans Cobb, the so-called active interaction is to hope that the robot will talk when it talks, and if it doesn’t talk, it will not talk. Be a person who understands you Robot, the design of active interaction therefore becomes a very important content.
"A robot is a anthropomorphic device. Its sociality is very important. There is no need for voice between the robot and the robot, but the interaction with people requires voice communication. Voice interaction or active interaction of behavior, its boundary Where it is and where it is moderate, the industry has paid more attention to it in recent years," said Jans Cobb.
The difficulty lies in this. "What makes it difficult for robots to complete complex tasks? A large part of the reason lies in the complex dynamics, that is, how the robot interacts with the surrounding environment. There will be many uncertainties and changes in the interaction with objects, environments and tasks. When interacting with humans, the situation may be more complicated." Jans Cobb said.
So how does a robot learn a new skill? Jens Cobb said that generally speaking, robots learn new things in two different but complementary ways. The first is imitation learning, that is, the teacher demonstrates a skill, and then the students try to imitate. This method is suitable for simple learning, but for more complex tasks, robots need reinforcement learning. Reinforcement learning is a kind of The goal of the learning method based on “wrong” approach is to allow artificial intelligence to take actions that maximize returns in a specific environment. If the machine is doing the right thing in the learning process, it will be rewarded accordingly. If it is wrong, there will be no reward. This learning method allows the machine to understand what is the right behavior and what is the wrong behavior in the interaction with the environment, instead of Use big data to make machines "rote". Jens Cobb emphasized that reinforcement learning allows machines to speed up the learning process and thus can complete more complex tasks. For humans, teaching in this way is also quite intuitive. This is also a feasible path for future robots to strengthen their learning and interaction capabilities.