Vision Based Lunar/Mars Rover Slip Prediciton Research

Wheeled robots commonly undergo slip in various types of natural terrains,particularly in planetary exploration missions.Predicting slip from a distance is important for safe driving,path planning,accurate localization and navigation.In recent decades,a large number of studies have been conducted to model slip from a mechanical point of view.However,because many factors that affect the slip,such as movement state,soil parameters,wheel characteristics,weight distribution of the vehicle body,and local topography,it is difficult to establish a physical model of the rover slip.For the same reason,even if a physical model is established,it is difficult to be generalized in other situations and more difficult to use for slip prediction.About a decade ago,in order to resolve the slip problem of NASA’s second generation Mars rover Opportunity and Spirit,a new computer vision and machine learning based slip prediction framework was proposed.In this framework,the factors affecting slip are decomposed into physical factors and geometric factors.The former is limited by the terrain type information provided by the camera image,and the latter is provided by the terrain geometric 3D reconstruction of the stereo camera.Comprehensive analysis shows that this is the only feasible method that can be used for slip prediction.However,due to less research in this area,there are still many defects so far.Therefore,this article focuses on three main issues,including the construction of the overall prediction framework,the geometric information based slip modeling,and navigation camera images based Mars terrain type recognition.The specific research contents and innovations are summarized as follows.(1)The overall framework of slip prediction based on computer vision and machine learning is refined and reconstructed.The key technologies and theories involved in this method were analyzed and studied.And the issues of follow-up application related to slip prediction were designed.(2)The geometric information based slip prediction model is improved and its prediction potential is analyzed.It is generally believed that when a rover and terrain type are given,the slip is mainly determined by the terrain geometry that the rover passes through.In order to improve the expressive ability of geometry based slip model,three new influence factors were introduced by collecting and researching more precise and sophisticated data,which extended the expression ability of the model from static description to dynamic description,and increased the dimension of the model.In order to analyze the validity of the newly introduced factor,a series of comparative experiments were designed and verified based on two different algorithms.The experimental results show that the newly introduced impact factors are all valid,and the improved model has achieved significantly better results.The average mean absolute error of slip prediction is less than 5%.According to further analysis,the predictive ability of this model can reach an error of less than 2%.This is comparable to the slip measurement error in this paper and also illustrates the prediction limits of this model.(3)The possibility of using geometric information based slip models for similar terrain types was studied.The model was built based on the indoor experiment data of Chang’e-3 mission,and then it was tested and analyzed based on the on-orbit data of Chang’e-3 and Chang’e-4.The results show that although the experimental data used in model building and testing were acquired in different terrain types,and various experimental conditions were not the same,the prediction results still showed considerable prospects.The correlation between the predicted series of slips and the measured values exceeds 0.8,indicating that the geometry based slip model has a good ability to migrate between similar but different terrains.Theoretically,the basic framework of this paper requires that the geometry-based model must be used on the same terrain type,while this study illustrates the relationship between physical information and geometric information for the decision ability of the final result in the slip prediction process.This not only lays a foundation for the on-orbit application of earth training off-line geometric slip prediction model,but also provides a basis for the use of this method in the fact that terrain type classification itself has unavoidable ambiguity.These conclusions will provide an alternative way for follow-up research.(4)The Mars terrain classification based on navigation camera images was studied.Research based on the historical data of real Mars missions can avoid many problems such as traditional analog data is too different from reality.Combining the research on Mars terrain types in historical literature,the Mars terrain types are redefined and divided.At the same time,through the analysis of previous Mars missions and a large amount of data,it is believed that the terrain type of Mars is generally stable.Therefore,an off-line model based on existing Mars data can meet the recent Mars terrain type recognition requirements.These provide a basis for the use of offline models.According to the characteristics of this study,a specific terrain type learning and prediction method was designed.Experiments were conducted using classical methods in the fields of computer vision and machine learning,including texture-based methods and methods based on deep convolutional neural networks.Through experiments based on actual data,the classification accuracy and practicability of various methods are illustrated,and the problems and solutions that will be faced in practical application are proposed,which will provide necessary reference for China’s future Mars surface exploration mission.