The Optimized Physical Model for Real Rover Vehicle Jun young Kwak The Robotics Institute Carnegie Mellon University Pittsburgh PA 15213 junyoung kwak cs cmu edu Abstract This paper presents the way to select appropriate features for rover vehicle model and to reduce the errors differences between two different vehicle models The fundamental idea for selecting features is to compare the error rates when the specific group of features is removed and to choose the best case Based on selected features and learning algorithm I find the optimal coefficients and adjust the system properties using those values Results from a relevant simulation experiment provide foundations to support and illustrate the benefit of the devised ways The paper concludes with several promising directions for future research 1 In trod u ction Over the past few years rover vehicles for Mars exploration have been rapidly developed based on fundamental technologies However the measure and prediction of the parameter related to a wheeled ground robot while driving is still complicated High levels of slip and friction can be observed on certain terrains and the specific condition which can lead to significant errors of the vehicle s movement inability to reach its goals or in the worst case getting stuck without the possibility of recovery 1 The reduction of cost for analyzing the vehicle model and setting up some experimental procedures is also getting an important issue Current Mars rover vehicles have a lot of parameters and properties and we have to consider select relevant features in parameters to reduce the vehicles complexity and get the optimal solution for exploration and navigation on the rough terrain To safely perform tasks in unknown or complicated environments learning algorithm can be applied to navigation on the rough terrain using different vehicle models Using an online learning method directly learns the mapping between two different vehicle models and data through the experiment The system can be trained by simply driving through representative dataset The optimal vehicle model that encodes structure of the real rover platform represents the coefficients like suspension friction and max torque and similarity on the same terrain This paper provides the methods for selecting the optimal features in real dataset and reducing the execution errors between two different vehicle models using learning algorithm Doing so provides extra information to the algorithm that allows it to operate in more than just a naive manner using whole features in dataset These suggested techniques show improvements for the reducing whole error rates during learning and execution procedures 1 1 Related work Using structure to constrain problem A limitation of the constrained method is that predictions are made independently in each small patch of terrain without including any spatial context This can cause problems when there is ambiguous feature data 3 If the system considers each of these patches independently it will give the same ground estimate for both and get at least one of them wrong This work relaxes the strong assumption of independence between selected features through the inclusion of spatial correlations Learning to predict slip for rover Slip can be defined as the difference between the different vehicles estimated by several factors This work predicts the amount of slip an exploration rover by learning from examples of traversing similar terrain Learning from examples allows the system to adapt to unknown terrains rather than using fixed heuristics or predefined rules 2 This works also consider slip learning in a nonlinear regression framework They describe a general framework to learn the functional relationship between visual information and the measured slip using training examples 2 2 1 Approach Assumptions In this work I use two different simulators CMU Vortex simulator and NASA JPL ROAMs simulator See Figure 1 These two simulators have their own vehicle models but the other parts like path planning algorithm is the exactly same In this paper to simplify the problem I assume that the terrain condition and internal control part of each vehicle model are the same Based on this assumption I concentrate on finding factors making the execution errors between two simulators like tire friction slippery suspension and max torque For the path planning algorithm I used heuristically guided RRT hRRT 4 algorithm to collect data and to test the vehicle models Figure 1 Vortex and ROAMs simulator Figure 2 Collecting data using vortex simulator 2 2 Collecting data The data used in the paper was collected from two different simulators In this work there are two different dataset and one is obtained from Vortex simulator and another is obtained from ROAMs simulator To obtain the data I repeated generating the path and storing vehicle information on the planned path See Figure 2 After repeating 20 experiments I got 798 data and each data consists of 59 features These features represent the physical vehicle model and include yaw of body linear velocity angular velocity quaternion of each part curvature current car position and current planned position Features Current car position x y z Current planned position x y z Yaw Curvature Actual velocity Quaternion of body x y z w Quaternion of each wheel x y z w 4 Linear velocity of body x y z Linear velocity of each wheel x y z 4 Angular velocity of body x y z Angular velocity of each wheel x y z 4 2 3 Selecting the optimal features in dataset 2 3 1 Algorithm In this part I explain the way to find optimal features or parameters in dataset with AdaBoost algorithm See Figure 3 a For selecting features I use AdaBoost algorithm because that makes the error low and help to reduce the accidental cases This algorithm provides the method for improving the accuracy of any learning algorithm and uses weak algorithms for single rules It also combines weak rules into a strong learning algorithm by weighting weak learners To apply AdaBoost algorithm to this problem I used linear classifiers Na ve Bayes classifier as a weak learner Figure 3 Used learning algorithm I use the physical vehicle information as features and difference between planned positions and current vehicle positions as the target information 1 Features X The physical vehicle model 59 features Yaw of body linear velocity angular velocity quaternion of each part curvature current car position and current planned
View Full Document
Unlocking...