Yamakawa Laboratory, The University of Tokyo

東京大学　山川研究室

Enhancing robotic performance on dynamic tasks is crucial for deploying robots in diverse real-world and human?robot collaboration (HRC). In this work, we focus on a dynamic benchmark task: robotic catching. To improve performance, we propose a coarse-to-fine catching strategy that combines coarse pre-throw catching pose estimation using Graph Convolutional BiLSTM (GC-BiLSTM) based throwing zone estimation from the pitching motion, followed by fine post-throw refinement of the catching pose based on the object's flight trajectory. Furthermore, we employ global?local skeleton joint motion decoupling and human height scaling to enhance the model's ability to capture spatiotemporal information effectively. These components are integrated into an end-to-end system, including object tracking and robot control, whose effectiveness we demonstrate in real-world experiments. Through extensive evaluations for throwing zone classification with six state-of-the-art models, the proposed GCBiLSTM consistently achieves the highest precision and F1 score, demonstrating strong classification performance even on unseen data while maintaining real-time computational capability, with an average inference time of 1.37 ms. Notably, the model achieves the highest F1 score of 75.4 % when trained on data from a single subject, compared with 61.7 % when trained on the mixed training dataset. Overall, we believe that our system provides a key approach for enhancing the robot's ability to handle dynamic situations and serves as a valuable reference for future skeletonbased action recognition research.