Fall Prediction for New Sequences of Motions

Motions reinforce meanings in human-robot communication, when they are relevant and initiated at the right times. Given a task of using motions for an autonomous humanoid robot to communicate, different sequences of relevant motions are generated from the motion library. Each motion in the motion library is stable, but a sequence may cause the robot to be unstable and fall. We are interested in predicting if a sequence of motions will result in a fall, without executing the sequence on the robot. We contribute a novel algorithm, ProFeaSM, that uses only body angles collected during the execution of single motions and interpolations between pairs of motions, to predict whether a sequence will cause the robot to fall. We demonstrate the efficacy of ProFeaSM on the NAO humanoid robot in a real-time simulator, Webots, and on a real NAO and explore the trade-off between precision and recall.