Force-based Real-time Inhomogeneity Detection for Robotic Uniform Mixing of Heterogeneous Materials

This study proposes a time-series learning approach for the real-time detection of inhomogeneity (clumping) in heterogeneous materials during robotic mixing using force-torque sensor data. Many existing robotic mixing systems execute predefined motion patterns that do not fully leverage dynamic changes in material states, resulting in limited responsiveness to transient mixing conditions. To better capture these dynamic variations, time-series force and torque data collected from a 6-axis sensor mounted on a robot arm are segmented into fixed-length temporal windows and used as model inputs. Ground-truth labels are established through offline analysis of synchronized external camera recordings. Three time-series classification models, namely a Bidirectional LSTM (BiLSTM), a Unidirectional LSTM (UniLSTM), and a 1D Convolutional Neural Network (1D-CNN), are trained and evaluated against a rule-based baseline. Experiments conducted across 43 mixing trials with proxy materials and evaluated using leave-one-out cross-validation show that all three learning-based models substantially outperform the rule-based approach, with BiLSTM achieving an average recall of 0.787, an F1-score of 0.704, and an AUC of 0.857. These findings indicate that leveraging temporal patterns in force- torque signals is more effective for clumping detection than approaches based solely on instantaneous signal values. Overall, this study demonstrates the potential of time-series learning to enhance mixing uniformity and support adaptive control strategies in autonomous robotic mixing systems.