Reproduce “Catching Spinning Table Tennis Balls in Simulation with End-to-End Curriculum Reinforcement Learning” on OpenResearch

About this paper

Xiaoyi Hu, Yue Mao, Gang Wang, Qingdu Li, Jianwei Zhang, Yunfeng Ji

A robotic agent was successfully trained in simulation to catch and return high-spin table tennis balls, leveraging an end-to-end curriculum reinforcement learning framework and a physics-based collision model. This approach achieved an 81% return rate with a 0.31m target error during Real2Sim validation using real-world ball trajectories.

Key insights

• Combining curriculum reinforcement learning with an accurate spinning ball collision model is essential for achieving optimal robot performance in highly dynamic, spin-dependent tasks.
• Defining distinct trajectory states and designing a stage-based reward function effectively guides learning for complex, cyclical tasks, allowing robots to acquire fundamental skills progressively.
• High-fidelity simulation, incorporating accurate physics models and domain randomization (e.g., perception noise), is crucial for robust Sim2Real transfer and for pre-assessing robot capabilities against real-world complexities.

Summary by alphaXiv.