Abstract:
Background: Robotic table tennis systems offer an ideal platform for pushing camera-based robotic manipulation systems to the limit. The unique challenge arises from the fast-paced play and the wide variation in spin and speed between strokes. The range of scenarios under which existing table tennis robots are able to operate is, however, limited, requiring slow play with low rotational velocity of the ball (spin).
Research Goal: We aim to develop a table tennis robot system with learning capabilities able to handle spin against a human opponent.
Methods: The robot system presented in this thesis consists of six components: ball position detection, ball spin detection, ball trajectory prediction, stroke parameter suggestion, robot trajectory generation, and robot control.
For ball detection, the camera images pass through a conventional image processing pipeline. The ball’s 3D positions are determined using iterative triangulation and these are then used to estimate the current ball state (position and velocity).
We propose three methods for estimating the spin. The first two methods estimate spin by analyzing the movement of the logo printed on the ball on high-resolution images using either conventional computer vision or convolutional neural networks. The final approach involves analyzing the trajectory of the ball using Magnus force fitting. Once the ball’s position, velocity, and spin are known, the future trajectory is predicted by forward-solving a physical ball model involving gravitational, drag, and Magnus forces.
With the predicted ball state at hitting time as state input, we train a reinforcement learning algorithm to suggest the racket state at hitting time (stroke parameter). We use the Reflexxes library to generate a robot trajectory to achieve the suggested racket state.
Results: Quantitative evaluation showed that all system components achieve results as good as or better than comparable robots. Regarding the research goal of this thesis, the robot was able to
- maintain stable counter-hitting rallies of up to 60 balls with a human player,
- return balls with different spin types (topspin and backspin) in the same rally,
- learn multiple table tennis drills in just 200 strokes or fewer.
Conclusion: Our spin detection system and reinforcement learning-based stroke parameter suggestion introduce significant algorithmic novelties. In contrast to previous work, our robot succeeds in more difficult spin scenarios and drills.