Proprioceptive localilzatilon foraquadrupedal robot onknownterra'in
Sachin Chitta · 2007
Wepresent anovel methodforthelocalization ofa legged robot onknownterrain using onlyproprioceptive sensors suchasjoint encoders andaninertial measurement unit. In contrast toother proprioceptive poseestimation techniques, this imethod allows forglobal localization (i.e., localization withlarge initial uncertainty) without theuseofexteroceptive sensors. Thisismadepossible byestablishing a measurement model basedonthefeasibility ofputative poses onknownterrain given observed joint angles andattitude measurements. Results are shownthatdemonstrate thatthemethodperforms better than dead-reckoning, andisalsoabletoperform global localization fromlarge initial uncertainty. I.INTRODUCTION Theproblem oflocalization using visual features hasbeen widely studied inrecent years. A typical approach istouse aknownmapoftheenvironment andupdate poseonline using acombination ofodometry andobservation ofknown features inthescene. Inanother approach, theprocess of localization andchoosing theright setoffeatures iscarried outsimultaneously giving rise totheapproach knownas SLAM.Inallthiswork,localization relies heavily onthe useofanexteroceptive sensor, suchasacamera, GPSunit orlaser scanner. Proprioceptive sensors, suchasjoint en- coders andangular rategyros provide odometry information whileaccelerometers canbeused(atrest) todetermine thedirection ofgravity. Verylittle workhasaddressed the problem oflocalization inknownenvironments using only proprioceptive sensing. Inthis work, wewill address theproblem oflocalization of aquadruped robot using only proprioceptive sensing. A good analogy tomotivate andexplain this concept isthat ofhuman motion inadarkbutfamiliar room.Initially, humanshave nosenseofwheretheyareintheroombuttaking acouple ofsteps andencountering, forexample, afamiliar feature like astep orrise instantly localizes usintheroomThereis still anelement ofuncertainty intheestimate ofposltion intheroom, forexample itmaybedifficult tolocalize along tbelength ofthestep. However, eventbelimited pose information gleaned fromthis dataisoften enough tomake meaningful decisions. Thistechnique, that wewill henceforth refer toasproprioceptive locali ationcanalsoprove very useful insituations wheretheprimary external sensors ofa robot fail. We arespecifically motivated bytheapplication ofthis technique tolegged robots. Legged robots cansenserich ground features and, ifappropriate sensors areavailable, also (a)Feasible configu- (b)Infeasible config- (c)Feasible configu- ration uration ration Fig.1.Illustration ofhowproprioception might beusedtodistinguish valid poses frominvalid ones. Fig.I(a) illustrates abodyposethat isconsistent withproprioception, given theassumption ofstatic stability; thezeropitch ofthebody isconsistent withalllegsbeing extended onaflat surface. Fig.I(b)illustrates abodyposethat isinconsistent withproprioception andthegiven terrain; since thefront legshavenothing torestoninthat configuration, wewouldexpect therobot tobepitched forward wereitin