Towards Robust Robotic Throwing: A Hybrid Framework of Physics and Learning
Robotic throwing extends a manipulator’s functional workspace by leveraging object flight kinematics,yet accurately hitting distant, cluttered targets is notoriously sensitive to grasp pose, release velocity, aerodynamics, and real-world model error. Traditional projectile equations ignore these complexities, while purely data-driven policies demand vast experience and often fail out of distribution. This thesis proposes a hybrid framework that fuses analytic physics with learning to achieve fast, precise, and generalizable robotic throwing. The system, implemented on a UR10 arm with an overhead Intel RealSense D455i camera, first predicts stable top-down grasps from RGB-D heightmaps. It then computes an analytic release velocity from closed-form projectile equations and refines that estimate with two complementary learning modules: a Residual-Physics Network that directly corrects model bias, and a Physics-Informed Neural Network (PINN) that embeds the governing differential equations into its loss for data-efficient trajectory prediction. All components are trained end-to-end with prioritized experience replay in both photo-realistic PyBullet simulation and on a 10-object YCB-based physical benchmark spanning diverse masses, shapes, and textures. Extensive experiments compare four settings namely Physics Only, Learning Only, Residual Physics, and PINN augmented control. The hybrid methods cut landing error by up to one-third relative to pure baselines, maintain high grasp success (> 78 %), and transfer to previously unseen objects and target bins without retraining. Ablations reveal that residual corrections handle object-specific aerodynamic effects, while PINN regularization improves sample efficiency and stabilizes training. These results demonstrate that coupling light-weight analytic models with structure-aware learning combines the interpretability and safety of physics with the adaptability of data, enabling reliable robotic throwing in unstructured environments. Beyond tossing, the framework is readily extensible to other dynamic skills—such as batting, sliding, or kicking—where accurate long-horizon physics reasoning is essential.
Advisor: Professor Constantinos Chamzas (WPI)
Committee: Professor Berk Calli (WPI) and Professor Mahdi Agheli (WPI)
Zoom: Zoom meeting has ended.