Fall analysis and prediction for humanoids

In bipedal locomotion, avoiding falls is a significant and challenging issue, and addressing this challenge requires precise fall prediction methods. Since humanoid robots’ state data is a temporal signal in a high-dimensional feature space, how to extract appropriate features and enable rapid real-time response in actual systems constitutes a significant challenge. We designed an algorithm of convex hull vertexes selection (CHVS) for fall analysis, selecting features that contribute more to assessing the robot's state. For the fall prediction, we first used an 11-dimensional statistical measure for feature extraction from the temporal signal. After state classification, we further incorporated the temporal accumulation effect, achieving a trade-off between minimizing false positive rates and maximizing advance prediction time. Subsequently, we designed a simple adjustment strategy for imminent falls and integrated it into the overall control framework. We validated the proposed fall prediction algorithm using the data of robot Q1. The validation was carried out under the condition of small sample training. We compared our algorithm with mainstream learning-based fall state classification methods, such as Long Short Term Memory Networks (LSTM) and Convolutional Neural Networks (CNN). The results show that the accuracy of state classification has increased by 3%–5%. At the same time, the computation time has been reduced by more than half. After time accumulation, the algorithm could predict more than 1 s in advance with a false positive rate of 0. The control framework incorporating the fall adjustment was tested in the simulation of the Cassie robot with an upper body. Some disturbances that could not be overcome under the original framework were adjusted to a normal motion state, proving the usefulness and feasibility of the proposed fall prediction for real-time dynamic scenarios to prevent falls.