Abstract— Myoelectric control of rehabilitation devices engages active recruitment of muscles for motor task accomplishment, which has been proven to be essential in motor rehabilitation. Unfortunately, most electromyographic (EMG) activity-based controls are limited to one single degree- of-freedom (DoF), not permitting multi-joint functional tasks. On the other hand, discrete EMG-triggered approaches fail to provide continuous feedback about muscle recruitment during movement. For such purposes, myoelectric interfaces for continuous recognition of functional movements are necessary. Here we recorded EMG activity using 5 bipolar electrodes placed on the upper-arm in 8 healthy participants while they performed reaching movements in 8 different directions. A pseudo on-line system was developed to continuously predict movement intention and attempted arm direction. We evaluated two hierarchical classification approaches. Movement intention detection triggered different movement direction classifiers (4 or 8 classes) that were trained and tested over a 5-fold cross validation. We also investigated the effect of 3 different window lengths to extract EMG features on classification. We obtained classification accuracies above 70% for both hierarchical approaches. These results highlight the viability of classifying online 8 upper-arm different directions using surface EMG activity of 5 muscles and represent a first step towards an online EMG-based control for rehabilitation devices. I. INTRODUCTION Rehabilitation devices such as robotic exoskeletons have shown great potential in the field of motor rehabilitation as they permit a repeatable and intensive proprioceptive stimulation of paralyzed limbs in terms of goal-oriented mobilizations. However, robot-aided treatments do not necessarily contribute to regain motor function unless active and voluntary participation of patients’ paretic muscles is present during movement [1], [2]. In this context, myoelectric interfaces for controlling robotic exoskeletons constitute a promising potential tool to effectively involve muscle recruitment for motor task accomplishment. Current myoelectrical interfaces applied in rehabilitation chiefly focus on the control of single-DoF actuators based on 1 N.I-L., A.S-S., F.S., E.L-L., C.B., N.B. and A.R-M. are with the Institute of Medical Psychology and Behavioral Neurobiology, University of Tübingen, Germany (corresponding author: nerea.irastorza- landa@medizin.uni-tuebingen.de) 2 N.I-L., A.S-S. and F.S. are with the IMPRS for Cognitive and Systems Neuroscience, Tubingen, Germany. 3 N.I-L., is with IKERBASQUE, Basque Foundation for Science, Bilbao, Spain. 4 N.B. is with the Wyss Center, Geneve, Switzerland. 5 amos-urguialday is with the eurotechnology aoratory an eastian pain EMG amplitude. EMG-based control systems that continuously modulate mechanical assistance provided by a robotic exoskeleton have been developed for single-DoF movements such as wrist extension [2] and hand grasping [3]. Nevertheless, while continuous control of a single DoF based on EMG activity can be feasible, the decoding of myoelectric signals during movements that involve several DoFs simultaneously remains still challenging, especially in stroke patients. Simpler approaches based on EMG-triggered interfaces have also been applied in order to train multi-joint tasks such as multi-directional movements with the upper arm [4]. This approach enables to trigger the movement of the actuator without the need of self-producing any actual movement, which may allow even highly-impaired participants to activate robot assistance [5]. Nonetheless, although EMG-triggered approaches might reinforce muscle strength by requiring constant activation of specific muscles above a certain threshold during movement, they fail to provide specific feedack regarding the user’s inappropriately co-activated EMG patterns. On the other hand, discrete approaches in which rehabilitation devices are triggered by EMG activation only at the trial onset fail to provide feedback about the muscle recruitment continuously during the entire movement execution. In spite of inappropriate co-activation of muscles due to incorrect muscle synergy recruitment in stroke patients [3]– [6], successful results in classifying residual EMG activity of stroke patients have been reported [7],[8]. Diverse functional movements involving multiple joints at the upper-limb, wrist and hand levels have been discriminated based on the residual activity of stroke patients, achieving better results in moderately impaired patients (71.3%) than in severely paralyzed subjects (37.9%) [7]. From these results, it is concluded that the recognition of different discrete movements involving several upper-limb DoFs based on residual EMG can still be feasible. Therefore, myoelectric interfaces for the EMG continuous classification of discrete multiple-joint movements present a promising option for the training of functional movements using rehabilitation devices. In this work, we developed a “pseudo-online” system to predict movement intention and attempted direction during point-to- point reaching movements in the horizontal plane based on the EMG activity of five upper-limb muscles in healthy participants. II. METHODS A. Experimental Protocol Eight right-handed naïve healthy participants (5 males, age 25±2.74) were involved in the experiment after giving written informed consent to the procedures approved by the Design of Continuous EMG Classification approaches towards the Control of a Robotic Exoskeleton in Reaching Movements Nerea Irastorza-Landa 1,2,3 , Andrea Sarasola-Sanz 1,2 , Eduardo López-Larraz 1 , Carlos Bibián 1 , Farid Shiman 1,2 , Niels Birbaumer 1,4 and Ander Ramos-Murguialday 1,5