Meta Chief AI Scientist Yann LeCun predicts that by the end of this decade, artificial intelligence technology will usher in a further revolution. He said that current artificial intelligence systems are still limited and cannot yet realize the dream of home robots and fully autonomous cars. At the Queen Elizabeth Engineering Prize for engineering ceremony, he pointed out that major technological breakthroughs are still needed for AI to understand and interact with the physical world. AI still has a long way to go before it can catch up with humans or animals. He said that current AI is good at “manipulating language” but its understanding of the physical world is still very limited.
He said there are still many scientific and technological challenges ahead, and the limitations of current systems mean that AI may usher in a new revolution in the next three to five years. If we hope to eventually build home robots and fully autonomous cars, we must let the system understand the real world.
Yann LeCun is working on a new system that aims to help AI understand reality by building a model to predict the behavior of the physical world. Meanwhile, Yann LeCun’s peer, Yoshua Bengio, who is also a Queen Elizabeth Engineering Award winner, warned that the safety of AI technology still needs to be strengthened, and called on the upcoming Global AI Summit in Paris to focus on this issue. Bengio expressed the hope that world leaders can better understand the impact of what they are doing, not only the power they are creating, which may have positive or negative effects, but also the risks brought by this power.
Incompatibility Between Robots and the Environment
The collision between self-driving cars and pedestrians has triggered a heated discussion at the social level. It reflects the challenges that self-driving cars face in terms of human-machine relations, especially social acceptance, at the current level of technology. Self-driving cars can also be said to be a kind of robot. This is mainly because they have the core features of robots such as intelligent decision-making, autonomous learning, environmental adaptation and complex interaction. What is easily overlooked is that robots are AI with physical form. Since they are physical entities serving human activities, they must exist in three-dimensional physical space and in social and economic production relations. Therefore, the incompatibility between robots and the environment should not be ignored.
If we look at it from the perspective that self-driving cars are also robots, it also reflects the increasing closeness of human-machine interaction. Before robots, including self-driving cars, go out of the real physical space, they need to undergo a lot of pre-factory tests in simulators and laboratories. In terms of the complexity of training, robots are more difficult. For example, without including the direction vector, the humanoid robot has more than 40 joints. Self-driving cars are more regular, and their movements mainly include turning left, turning right, moving forward, moving backward, stopping, etc. If various complex environments and tasks are superimposed, the difficulty coefficient will be higher.
Environment Becomes an Important Factor in Promoting the Application of Robots
The problems faced by self-driving cars also remind us that the environmental dimension is also an important factor affecting the landing and application of robots in the future and becoming a new quality productivity. The environment should be three aspects, in addition to the social environment, it also includes the physical environment and the digital environment. The physical environment includes but is not limited to the physical space such as cities, buildings and roads where the robot operates, as well as the interactive environment of the robot hardware body. The digital environment includes computing, network environment, and also includes the simulation environment for robot training.
Just like cars need to be equipped with various radars, cameras and other sensors, the robot itself will also have many signal acquisition devices to improve the sensitivity of the perception system. With the maturity and commercialization of sensor technology, and the verification of robot operation rules and algorithms, sweeping and delivery robots have been able to enter life scenes. In this wave of technology for large models, AI also plays a very decisive role in training to improve the stability and response speed of decision-making algorithms. At the same time, the integration of soft and hard body control and even wearable electronic skin is also gradually breaking through.
At the planning level, robots need to be able to deeply understand complex logic and world knowledge, such as home environment, and iron plates cannot be placed in the microwave to heat up. At the control level, there are more simulation and real training. By setting rewards and penalties for safe reinforcement learning for robots, robots are constantly encouraged to avoid danger. In this process, the model is like a toddler, and the pain of falling allows children to learn to avoid dangerous actions. At the body level, the robot’s body also needs to be more evolved. Through reasonable mechanical limit design to limit the robot’s range of motion, force, torque and tactile sensor use, the robot can better perceive feedback.
The Main Challenges Faced by Robots in Digital Environments
The current simulation environment cannot fully simulate the real physical environment, and the simulation training of embodied intelligence focuses on the performance of the robot itself. At present, simulation training is widely used because it is already very economical. This requires a large amount of real and simulated data to feed robots for learning, training and evolution. The lack of data is one of the current engineering problems. But on the other hand, examples in training can never be exhaustive. At the intersections of traffic arteries in first-tier cities, the density of human-machine mixing is very high. This further requires robots to have the ability of evolutionary learning like humans, so as to deal with emergencies.
In addition to simulation training, there must be a high-performance computing device in the robot body that can process the models and rules on the end side. In addition, a digital environment such as a relatively high-speed and smooth network environment is also needed to cope with the problem that robots cannot call large models in the cloud. These are some of the main challenges and coping ideas faced by robots in the digital environment.
The goal of robots should be to better utilize physical rules, imitate human thinking, behavior and intelligence, and thus help humans solve problems. At present, when people discuss the drastic changes in the relationship between man and machine, they often ignore that robots are AI with entity tense. Since it is a physical entity serving human activities, it must exist in three-dimensional physical space and in social and economic production relations.
