Learning Discriminative Intention Predictors for Sit-to-Stand Assistance

Prediction of user motion intention is crucial for assistive robotics. Many previous studies have focused on only a single target motion. However, in practice, the user may also perform other motions. If a motion intention predictor constructed for the target motion is applied to other non-target motions, incorrect intention would be predicted and improper assistance might be provided to the user. To alleviate this issue, in this paper, we propose a discriminative intention predictor that has the capabilities of both intention prediction and target motion discrimination from non-target motions. Such a predictor can be constructed by learning latent features between the sensor input and prediction output of motion intention with the both target and non-target motion data. We applied it to the task of sit-to-stand timing prediction from surface electromyography and motion sensors. The experimental results show that our motion intention predictor certainly has capabilities of both intention prediction and motion discrimination. Furthermore, subjective experiments with a sit-to-stand assistance robot suggest that our approach can reduce the risk of improper assistance when the user performs non-target motions.