Specialist low-level controller for fast balls

Develop and train a low-level controller (LLC) specialized for returning fast incoming balls in the hierarchical robot table tennis system that uses an ABB IRB 1100 arm mounted on Festo linear gantries, overcoming the data scarcity and hardware limitations that previously prevented successful training of such a specialist.

Background

The system trains a library of low-level controllers (LLCs) to handle specific table tennis skills such as forehand/backhand play, targeting, and serving. During specialization, the authors attempted to create LLCs for various scenarios, including fast balls, lobs, and spin-specific serves. Despite building a comprehensive dataset and iterative sim-to-real cycles, they report a specific inability to train an LLC specialized for fast balls due to insufficient data and hardware constraints.

Addressing this limitation would directly improve the robot’s performance against higher-speed opponents and reduce failure modes linked to system latency and reset constraints noted elsewhere in the paper. A successful fast-ball specialist would complement targeting and spin-specialized LLCs, enhancing the overall adaptability and competitiveness of the high-level controller (HLC).