Visual–tactile learning of robotic cable-in-duct installation skills

Abstract

Cable-in-duct installation is one of the most challenging contact-rich interior finishing tasks for construction robots. Such precise robotic cable manipulation skills are expected to be endowed with high adaptability towards unstructured on-site construction activities via Sim2Real transfer. This paper presents a Sim2Real transferable reinforcement learning (RL) policy learning method for multi-stage robotic cable-in-duct installation, employing reward shaping to support unified task completion through a multi-stage RL policy. Specifically, the Foreground-aware Siamese Tactile Regression Network (FSTR-Net) is introduced as a feature-level unsupervised domain adaptation method to enhance the Sim2Real transfer of the RL strategy. Evaluations demonstrate that the robotic skill for cable-in-duct installation attains a success rate exceeding 98% in the simulator. FSTR-Net achieves over 99% accuracy for tactile-based in-hand fish tape pose estimation. Furthermore, real-world experiments show an average success rate of 95.8%, validating the RL strategy’s generalization and the approach’s effectiveness in mitigating the domain gap.