Feature-Space Reinforcement Learning for Robotic Manipulation

Reinforcement Learning (RL) has gained popularity for developing intelligent robots, but challenges such as sample inefficiency and lack of generalization persist. The choice of observation space significantly influences RL algorithms' sample efficiency in robotics. While end-to-end learning has been emphasized, it increases complexity and inefficiency as the agent must re-learn forward and inverse kinematics. To address these issues, we propose a straightforward approach that utilizes readily available control techniques, such as forward and inverse kinematics, to capitalize on domain knowledge. Our approach involves enhancing the observation space with task-space features and utilizing task-space inverse kinematics. Our contributions include a proposal for mathematical formulation and a framework for RL algorithms in robotics.