During hands-on robotic surgery it is advisable to know how and when to provide the surgeon with different assistance levels with respect to the current performed activity. Gesteme-based on-line classification requires the definition of a complete set of primitives and the observation of large signal percentage. In this work an on-line, gesteme-free activity recognition method is addressed. The algorithm models the guidance forces and the resulting trajectory of the manipulator with 26 low-level components of a Gaussian Mixture Model (GMM). Temporal switching among the components is modeled with a Hidden Markov Model (HMM). Tests are performed in a simplified scenario over a pool of 5 non-surgeon users. Classification accuracy resulted higher than 89% after the observation of a 300 ms-long signal. Future work will address the use of the current detected activity to on-line trigger different strategies to control the manipulator and adapt the level of assistance.