Three stages of the real-time classification experiments are explained and the addresses for supplementary visuals are provided.
The first experiment was conducted on a commercially available hand, Bebionic, for only a two-class (open and close) condition. In this experiment,
the state was updated at each 50 ms, and when the classification confidence is higher than the threshold (selected as 75% depending on the trade-off
between decision delay and correct gesture determination), the gesture decision was sent to the hand. The final finger positions were pre-defined
and the hand executed the final gesture by moving to these positions. Secondly, instead of an expensive and complex hand, another two-class classification
(pinch and medium wrap grasping) experiment was conducted on a 3D-printed hand. The same classification framework was repeated and an example visual is provided
in. Lastly, we investigated a more complex problem including three gestures (pinch, medium wrap and index pointing) and real-life object grasping experience
for demonstrating the suitability of the proposed system to real-life problems. In this experiment, while following the same methodology
the threshold was reduced to 50% confidence for considering the higher complexity of the problem due to higher number of possible classes.