Idsia & supsi & usi | continuous control behavior learning and adaptive robot operation based on Reinforcement Learning

【 title 】Continuous control actions learning and adaptation for robotic manipulation through reinforcement learning

【 The author team 】Asad Ali Shahid, Dario Piga, Francesco Braghin, Loris Roveda

【 Date of publication 】2022.2.9

【 Thesis link 】https://link.springer.com/content/pdf/10.1007/s10514-022-10034-z.pdf

【 Recommended reasons 】 This paper presents a learning based approach , Using simulation data , Two model free reinforcement learning algorithms are used （RL） To learn object manipulation tasks . The learning performance of policy based and non policy based algorithms is compared ： Near end strategy optimization （PPO） And soft participant criticism （SAC）. In order to speed up the learning process , A fine tuning program is proposed , This program demonstrates policy based RL Continuous adaptation to the new environment , Allow the learned strategies to adapt and execute （ part ） Modified tasks . A dense reward function is designed for the task , In order to realize the effective learning of agents . Will involve Franka Emika Panda The grasping task of the manipulator is regarded as a reference task to learn . The learned control strategy can be extended to multiple object geometries and initial robots / In part configuration .