Summary of Key Points
Recently, humanoid robots have made significant advancements in racket sports such as table tennis and tennis: they are not only capable of playing autonomously against Olympic champions (Ding Ning, Zheng Jie) but have also participated in formal competitions (e.g., the World Human Robot Games Table Tennis Tournament). Racket sports have become a new arena for robotics competition, as they test the complete “perception-decision-execution” cycle. These capabilities can be applied to real-world scenarios such as industrial sorting and domestic services. Currently, the robots’ performance is comparable to that of human beginners, but they still face challenges in capturing fast-moving balls, coordinating decision-making and execution, and controlling costs. Commercialization is still in the testing phase, but it is expected that robots will be available as training companions in the next 2-3 years and may eventually enter households or professional sports training.
1. Racket Sports: The “New College Entrance Exam” for Robots
Why do robots practice table tennis? Because it is a true test of their technology. The founder of Qingtong Vision categorizes robot sports into three types:
- Routine Sports (e.g., dancing): These can be programmed in advance, and the technology is mature.
- Racket Sports (e.g., table tennis, tennis): These require real-time perception of the ball’s trajectory, quick decision-making on how to hit the ball, and coordination of body movements; initial progress has been made.
- Team Sports (e.g., football, basketball): These involve collaboration and strategy, and are still in the exploration stage.
Racket sports are like a “college entrance exam” for robots; if they pass this test, it indicates that their “perception (eyes), decision-making (brain), and execution (body)” capabilities are up to standard. These skills are highly relevant to real-world applications, such as picking up trash (identifying the trash, planning the movement, and accurately picking it up).
2. Why is it So Difficult for Robots to Play Sports?
For humans, racket sports are instinctive, but for robots, it represents an “epic engineering challenge”:
- Perception Difficulty: A table tennis ball travels from serve to cross the net in just 0.15 seconds (faster than a blink of an eye) at a speed of 15 meters per second, and it may be spinning. Traditional cameras often blur or lag, so multiple high-speed cameras (capable of millisecond-level positioning) are used during competitions, and high-frequency vision algorithms (capturing 20,000 times per second) are required for accurate ball tracking during training.
- Decision-Making and Execution Difficulty: The robot’s “brain” and “cerebellum” must work together. The brain must determine where the ball is going and how to hit it in milliseconds, while the cerebellum translates this into bodily actions (e.g., adjusting steps and the angle of the swing). In this competition, all teams used the Zhiyuan robot platform, which tested the algorithms’ ability to coordinate these processes efficiently.
- Hardware Limitations: Robots need to be lightweight yet powerful (the Zhiyuan Yuanzheng A3 is 1.73 meters tall and weighs only 55 kilograms, with joint torque of 400 Nm); otherwise, they lack speed and explosive power. However, being too light can lead to instability.
2. Different Teams’ Approaches to Solving These Challenges
Two technical approaches have been adopted by different teams:
- Layered Architecture (HKU, Berkeley): This is similar to a company’s division of labor, with a higher-level “brain” responsible for calculating the ball’s trajectory and hitting strategies, and a lower-level “cerebellum” using reinforcement learning to control movements. HKU also collected data from human athletes for two months to train the robots to perform advanced techniques like lifting and smashing the ball.
- End-to-End Architecture (Tsinghua University): This approach is more radical, allowing the robot to go from “seeing the ball” to “taking action” directly without intermediate steps. The advantage is that it enables the robot to learn complex techniques like backhand smashes, but the training is more difficult (it’s like teaching the robot to learn on its own).
The HKU team plans to evolve towards an end-to-end architecture in the future to achieve even better performance.
3. Commercialization: From Experience Centers to Professional Training Companions
Robots playing racket sports are still in the testing phase:
- High Costs: High-speed cameras cost tens of thousands of yuan each, and a complete system requires several cameras, making large-scale deployment unfeasible. HKU is exploring a “pure vision” approach that does not rely on external cameras, using the robot’s own cameras. They have tested this approach outdoors and found that it is attractive to the public. In the professional sports context, Zheng Jie hopes that robots can simulate Serena Williams’ serves to provide customized training for athletes.
- Time to Market: HKU estimates that robots will reach a training companion level in table tennis within 2-3 years, while Galaxy General is more optimistic, predicting that they will reach expert levels in tennis in 1-2 years and potentially challenge world champions.
4. Beyond Racket Sports: Where Else Can These Technologies Be Applied?
The skills acquired from racket sports training are not wasted:
- Industrial Applications: For example, in sorting fast-moving items, robots need to quickly perceive, decide, and act.
- Domestic Services: Robots can help with tasks like picking up trash, requiring precise positioning, movement planning, and accurate execution.
- Medical Rehabilitation: Robots can assist in rehabilitation training by precisely controlling the intensity and angle of movements.
Racket sports serve as a practical training ground for robots; mastering these skills brings them one step closer to being integrated into factories and households.
In summary, the advancements in humanoid robots’ racket sports skills are not just for fun but are aimed at developing core technologies for real-world applications. Although there are still many challenges, the robots are learning rapidly. Who knows? In a few years, you might have a robot playing table tennis with you at home.