Skill Learning for Manipulator using Residual Policies and PINNS
Date:
Developed Residual Policies: Designed a learning framework combining model-based planning with reinforcement learning to improve adaptability in robotic throwing tasks. Implemented PINNs: Integrated Physics-Informed Neural Networks to embed domain knowledge into the learning process, ensuring accurate modeling of physical constraints such as dynamics and kinematics in the throwing task. Skill Learning for Robot Throwing: Focused on precise trajectory planning and execution by leveraging hybrid approaches that combined data-driven learning with physics-based models. Simulation Environment Design: Created a realistic simulation environment in PyBullet to train and evaluate throwing tasks under controlled conditions. Optimized Training Processes: Enhanced efficiency and accuracy in skill acquisition by integrating PINNs and Residual Policies, leading to improved performance in robotic throwing. Real-World Relevance: Bridged the gap between simulation and real-world deployment by ensuring the learned policies were robust and transferable.