Learning-driven sensorless interaction force estimation for low-cost robot arm with limited dynamic features

Abstract

Precise measurement of the interaction force between the robot and its environment benefits the decision-making processes in various robotic applications. Compared with sensor-based methods, sensorless approaches are commonly preferred due to their versatility and cost-effectiveness. This paper introduces a learning-based method that leverages the state-of-the-art transformer to accurately estimate the interaction force. In contrast to other estimation methods relying on accurate robot dynamic parameters, state information or image features, a notable innovation of our work is the utilization of the limited set of features. The elaborate feature set only includes the joint angle, velocity, and driven torque, with the omission of joint acceleration—a basic robot state typically employed in other research. This configuration expands the feasibility of the presented approach to low-cost robots which are solely equipped with encoders in each joint, and to scenarios where the collection of clear and unobstructed visual features are challenging. Another distinctive feature of our work is that both soft and stiff objects during interaction are considered. Results from the experiment demonstrate that, in comparison to previous image-based methods, our framework achieves an equivalent or even superior level of accuracy across a broader spectrum of environments. Additionally, due to the elimination of joint acceleration from the feature set, the proposed framework sacrifices a small degree of accuracy compared with some non-image-based methods to broaden its applicability.