V. Sivtsov et al.: Intrinsic Reward Decomposition for Soft Robotic Manipulation Tasks скачать в хорошем качестве

V. Sivtsov et al.: Intrinsic Reward Decomposition for Soft Robotic Manipulation Tasks

   • 2025 IEEE ICDL Prague   — https://icdl2025.fel.cvut.cz — 2025 IEEE International Conference on Development and Learning (ICDL) Prague ●● ► Vladimir Sivtsov, Daniil Shkolnik, Athanasios Papanikolaou, Ivan Markovic, Ivan Petrovic, Enrica Zereik, Fabio Bonsignorio: Intrinsic Reward Decomposition for Soft Robotic Manipulation Tasks ●● Reinforcement learning has shown remarkablepotential for autonomous skill acquisition. However, effectiveexploration-exploitation of possible actions and states remainsa fundamental challenge, especially in soft robotic manipulationtasks where the continuous deformation of soft materialscauses complex nonlinear dynamics and high-dimensional statespaces. A promising solution to this problem is intrinsicmotivation, which allows learning agents to explore the environment more systematically by producing self-supervisedreward signals. Existing intrinsic reward techniques, however,frequently approach the whole state space as a single entity,which may reduce their efficacy in robotic manipulation, whereinteractions can take place between the manipulated itemand the robot. In this work, we introduce a new method ofintrinsic reward decomposition which focuses exploration ontask-relevant interactions. Our method implements a weightedcombination of random network distillation rewards derivedseparately from robot observations and manipulation objectstates. Experimental results across various manipulation tasksin the soft robotic benchmark show that this attention-inspireddecomposition enables more effective discovery of manipulationstrategies and significantly enhances performance.