New Challenges for Robots in Achieving Motion Intelligence
In the field of artificial intelligence, the famous Turing test proposed by British mathematician Alan Turing in 1950 is used to determine if a computer program possesses the same intelligence as a human being. Similar tests exist in the field of universal footed robots, where the ability to grab any cup under any circumstances is considered as true motion intelligence. However, passing this test is extremely difficult and requires robots to have strong perception capabilities and real-time interaction with the environment.
According to Zhang Wei, founder of Zhuji Dynamics, a general-purpose footed robot company, passing the test for footed robots is even more challenging than autonomous driving. While autonomous driving only requires positional relationship perception, legged robots need to add contact relationship perception, such as terrain and tactile perception. They must interact in real time with the environment to achieve motion intelligence. Unfortunately, most legged robots currently lack strong perception, limiting their mobility, efficiency, and robustness in complex scenes.
Zhang Wei emphasizes that the core of sports intelligence in robots comes from perception, specifically real-time perception based on complex scenes. To achieve precise and efficient control, robots must understand instructions from users and utilize environmental information to adjust motor movements. For example, a legged robot needs to know its relative position to stairs, terrain conditions, and how much force to use to navigate stairs. This dynamic perception-based motion control is challenging to achieve, and current legged robots face issues of efficiency and stability.
The structure of a robot’s legs is crucial for providing various services like a human, as it provides a higher degree of freedom. Legged robots must be able to adjust their center of gravity in real time and maintain balance to move in complex scenes, unlike their wheeled counterparts that struggle with unstructured terrains. Zhang Wei envisions two directions for future robot development: dedicated robots for single or simple scenarios, and multi-purpose robots that can handle complex terrains and tasks.
Although legged robots have made significant progress, there is still a long way to go before truly intelligent legged robot products become available on the market. The lack of perception is a major bottleneck for current quadruped robots, as it hinders their ability to dynamically sense and plan actions in real time based on the environment. Efficient and stable motion control requires the integration of positional relationship perception and contact perception, posing additional challenges.
To overcome these challenges, current robot control methods include model-based control and learning-based control. Learning-based control, which focuses on reinforcement learning, allows robots to iterate and improve their controllers through trial and error. However, perceptual interaction in complex scenes requires learning to fuse data from various sources.
In conclusion, achieving motion intelligence in legged robots is a complex and demanding task. The integration of real-time perception and motion control, along with efficiency and stability, are critical factors for the successful commercialization and implementation of legged robots in complex business scenarios. Only by addressing these challenges can legged robots provide true value and achieve a closed-loop business.