The article discusses the challenges of applying robotic manipulation in varied scenarios due to the limitations of control-based approaches and inefficiency of existing learning methods. To address these, the authors propose a novel method for skill learning in robotic manipulation: Tactile Active Inference Reinforcement Learning (Tactile-AIRL). This method aims to enhance the performance of reinforcement learning by integrating model-based techniques and intrinsic curiosity into the learning process, improving training efficiency and adaptability to sparse rewards. The use of a vision-based tactile sensor provides detailed perception for manipulation tasks, and a model-based approach is employed to imagine and plan appropriate actions. The method has shown high training efficiency in simulations and physical experiments, demonstrating its potential for practical applications.
Publication date: 21 Nov 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2311.11287
