求高手迅速英語(yǔ)翻譯成中文
whentherobotsenses,andwhenitmoves,respectively.
Suppose the robot just sensed s.Markov localization then
P (l j s) = ff P(s j l) P(l)
where ff is a normalizer that ensures that the resulting prob-
abilitiessumuptoone. Whentherobotmoves,Markov
localization updates P(l)
ability:
P ( 0l) =
using the Theorem of total prob-
Z
P ( 0lj a;l) P(l) dl
Here adenotesanactioncommand.Thesetwoupdate
equationsformthebasisofMarkovlocalization. Strictly
speaking, they are only applicable if the environment meets
aconditional independenceassumption(Markovassump-
tion), which specifiesthat the robot's pose is the only state
therein. Put differently, Markov localization applies only to
static environments.
Unfortunately, the standard Markov localization ap-
proachis prone to fail in denselypopulated environments,
sincethose violate the underlying Markovassumption.In
the museum, people frequently blocked the robot's sensors,
asillustrated in Figure 1.Figuratively speaking,if people
line up as a "wall" in front of the robot—which they often
did—,thebasicMarkovlocalizationparadigmmakesthe
robot eventually believe that it is indeed in front of a wall.
Toremedythisproblem,RHINOemploysan"entropy
filter" (Fox et al. 1998b). This filter, which is applied to all
proximity measurementsindividually, sortsmeasurements
intotwobuckets: onethatisassumedtocontainallcor-
rupted sensor readings, and one that is assumedto contain
onlyauthentic(non-corrupted) ones. Todeterminewhich
sensorreadingiscorruptedandwhichoneisnot,theen-
tropy filter measuresthe relative entropy of the belief state
before and after incorporating a proximity measurement:
P(l) logP(l) dl +P(l j s)logP(l j s) dl
lSensorreadingsthatincreasetherobot'scertainty
(_H(l;s) > 0) are assumed to be authentic. All other sen-
sor readings are assumedto be corrupted and are therefore
notincorporatedintotherobot'sbelief. Inthemuseum,
certainty filters reliably identified sensor readings that were
corruptedbythepresenceofpeople,aslongastherobot
knew its approximate pose.Unfortunately, the entropy fil-
ter canpreventrecoveryoncethe robot loosesitsposition
entirely.To prevent this problem, our approach also incor-
porates a small number of randomly chosen sensor readings
in addition to those selected by the entropy filter.See (Fox
et al. 1998b) for an alternative solution to this problem.