Please login first
Inclusive Human Intention Prediction with Wearable Sensors: Machine Learning Techniques for the Reaching Task Use Case
1 , * 1 , 2 , 2 , 1
1  University of Brescia
2  Institut Systèmes Intelligents et de Robotique (ISIR), Sorbonne Université



Predicting human intentions is a challenging task, which is gaining importance as well as collaborative robots (cobots) use and human-robot interaction, in several contexts, like industrial and clinical. The use of wearable sensors allows gathering data reducing instrumental or motion constraints, whereas machine learning techniques (MLT) can face the limit of small data amounts typical of these contexts. This study aims to compare MLT performance for the subject intention prediction in the illustrative scenario of a reaching movement, using three-dimensional position data gathered from wearable electromagnetic sensors.


Nine subjects (4 males; mean age: 51[years], range [29;71]) were recruited and asked to perform 3 repetitions of a sitting reaching movement for combinations of direction (left, center, right), quote (high, low), and distance (proximal, distal) [1], touching the goal and returning. An electromagnetic tracking system (Polhemus Fastrak) was used, with four sensors placed on acromion, upper third of humerus, wrist dorsum and manubrium. Data were elaborated in MATLAB environment to predict the goal position using only the information coming from a first sample of the acquired data. Linear Discriminant Analysis (LDA) and Random Forest (RF) algorithms were implemented, testing several sample dimensions.


LDA and RF prediction accuracy is computed and compared with respect to the data sample dimension. Considering a sample equal to 1/10 of the total movement (average time length t: 0.27[s]), LDA presents an accuracy of 81% (Standard Deviation SD 0.044), and RF of 73% (SD 0.012). Increasing the sample at 1/7 (t: 0.37[s]), accuracy rises at 89% (SD 0.034) with LDA and 83% (SD 0.011) with RF.


Both algorithms achieved good accuracy, which improves as the sample dimension increases. LDA presents better results. Both MLT give encouraging results and could be exploited in a collaborative scenario.


[1] J. V. G. Robertson et al. (2012), “Influence of the side of brain damage on postural upper-limb control including the scapula in stroke patients”.

Keywords: human intention prediction; wearable sensors; machine learning; reaching movement