Robot motor skill acquisition based on episodic memory and meta learning

Aiming at the problem of robots using past experience and knowledge to guide new task learning for few shot learning, the episodic memory was used to model the robot's experience, and an episodic-guided robot skill learning method was proposed based on meta-learning to achieve efficient learning of new tasks with few samples. The episodic memory was integrated to model the robot experience. During the learning process, the task experience passed by the meta-learner was stored as episodic memory through the encoder. In the process of guiding the learning of new tasks, the tasks were addressed hierarchically through the key value feature vector. The behaviors in the scene were extracted and combined until the task was completed. A meta-learning model was built based on a meta-learning network with long short term memory (LSTM) network as the core, and a normalization strategy and SeLU activation function were used to optimize the model to improve the robustness of the model. The UR3e robotic manipulator was used to train and test in simulation environment and real environment. The experimental results show that the proposed learning method has a success rate and good generalization in learning similar tasks with few shot learning. © 2021, Editorial Board of Journal of Huazhong University of Science and Technology. All right reserved.