Coarse-To-Fine Robot Catching Strategy With Gc-Bilstm Human Intent Recognition
Abstract
Enhancing robot 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: robot catching. To improve catching performance, we propose a coarse-to-fine catching strategy. Before the throw, the robot estimates the catching pose by predicting the throwing zone from the human pitching motion using a Graph Convolutional BiLSTM (GC-BiLSTM). After the throw, the estimated pose is refined using the object's flight trajectory. To further enhance spatiotemporal feature extraction, we introduce a rigid-body-dynamics-inspired skeleton representation that incorporates global-local joint motion decoupling and human height scaling. These components are integrated into an end-to-end system comprising throwing-zone prediction, object tracking, and robot control. Through extensive evaluations against six state-of-the-art models, the proposed GC-BiLSTM consistently achieves the highest precision and F1 score while maintaining real-time performance, with an average inference time of 1.37 ms. Notably, the model achieves its highest F1 score of 75.4% for a single participant, compared with 61.7% for six right-handed participants. Furthermore, a real-world robot catching experiment demonstrates that the proposed coarse-to-fine catching strategy enables earlier initiation of the catching motion, reducing the required catching speed by 35% compared with a conventional post-throw-refinement-only method. These results highlight the potential of the proposed system for dynamic robotic task execution and skeleton-based action recognition.