A Tactile-Driven Hierarchical Reinforcement Learning Framework for Dexterous Robotic Manipulation

Abstract

To address the challenges of robotic dexterous manipulation in complex environments, this paper proposes a hierarchical reinforcement learning (HRL) framework driven by tactile information. We design an adaptive hierarchical decision-making algorithm that integrates multimodal tactile features, dynamically adjusting hierarchical strategies and reward functions to adapt efficiently to complex tactile environments. Experiments were conducted on the PyBullet simulation platform, constructing a manipulation scenario involving 20 objects of varying shapes and materials. Two primary tasks were evaluated: basic grasping and placement, and complex assembly. Comparative results against standard DQN, PPO, and existing tactile-driven algorithms demonstrate that the proposed method achieves a success rate of 92.3% in basic tasks—outperforming DQN by 27.6% and PPO by 21.5%—while reducing the average operation time to 3.2 seconds. For complex tasks, the success rate increased to 85.7% (a 31.2% improvement over baselines), with a 40% acceleration in convergence speed. Ablation studies further validate that multimodal tactile feature fusion contributes an 18.3% increase in success rate, while the adaptive adjustment mechanism reduces strategy adjustment time by 35%. These findings confirm that the proposed framework significantly enhances robotic dexterous manipulation performance, offering a new pathway for the development of intelligent robotic systems.