Abstract
<title>Abstract</title> <p>Reliable force perception is essential for robots that physically interact with their environment, especially in applications involving manipulation, tool use, teleoperation, and contact-rich tasks. However, force/torque sensors mounted on industrial robots can produce biased measurements when unknown or interchangeable tools are attached to the end-effector. These biases become more difficult to compensate when the tool holder’s mass, center of mass, and inertial properties are not precisely known. This study presents a data-driven self-calibration framework for a KUKA KR 600 robot equipped with an ATI Gamma force/torque sensor to compensate force offsets under both quasi-static and dynamic conditions without requiring prior knowledge of tool properties. In the quasi-static case, a third-degree polynomial regression model was trained using end-effector orientation data to estimate gravity-induced offsets. The model achieved a test-set MSE of 0.005 N and a benchmark RMSE of 0.036 N. For dynamic calibration, more than 80,000 motion samples were collected during robot motion. After removing the quasi-static component, the remaining dynamic offset was modeled using polynomial regression and a multilayer perceptron (MLP). The second-degree polynomial model achieved a benchmark RMSE of 0.0458 N, while the MLP achieved a test-set RMSE of 0.041 N. These errors were below the practical accuracy limit of the ATI Gamma sensor, indicating that the proposed method can provide accurate force compensation under both static and dynamic conditions. The results show that the proposed framework improves force perception without manual tool modeling or hardware modification. Overall, by enabling robots to separate tool-induced disturbances from true interaction forces, this method supports more reliable robotic interaction, adaptive manipulation, and future integration into autonomous, haptic, and teleoperated robotic systems.</p>