Boundary-Uncertainty Active Learning for Efficient Segmentation of Robotic Performance Trajectories

Robotic art installations generate rich, high-frequency motion data during live performance, yet this data is rarely used by AI that could model, analyse or generatively reproduce the behaviours it captures. Performance recordings are unstructured, variable in length and arrive unlabelled, often requiring segmentation through manual segmentation that is prohibitive at scale. We present an active-learning pipeline that transforms raw recordings from Embrace Angels (2025-), an ongoing robotic art performative project in which a pair of Franka Panda arms embrace audience members, into a stage-segmented dataset suitable for future modelling. The pipeline combines an integrated annotation interface, self-supervised pretraining and a boundary-uncertainty acquisition function. We benchmark segmentation methods spanning zero-label baselines, classical template matching, window-based classical active learning with three base machine learning models and three neural sequence-model variants. A Temporal Convolutional Network (TCN) initialised from self-supervised pretraining achieves a boundary mean absolute error of 5.9 $\pm$ 0.9 seconds with a 76% reduction in annotation effort relative to fully manual labelling. Overall, active learning combined with self-supervised pretraining is particularly well suited to performance datasets where intra-class variability is high, class boundaries are inherently subjective and the marginal value of each additional labelled example can be extremely informative.