Efficient Spatial Concept Formation by Active Exploration of the Environment

Abstract

Autonomous service robots are required to adaptively learn the categories and names of various places through the exploration of the surrounding environment and interactions with users. In this study, we aim to realize the efficient learning of spatial concepts by autonomous active exploration with a mobile robot. Therefore, we propose an active learning algorithm that combines sequential Bayesian inference by a particle filter and position determination based on information-gain in probabilistic generative models. Our experiment shows that the proposed method can efficiently determine the position to form spatial concepts in simulated home environments.

Full text

؀ڥͷೳಈతͳ୳ࡧʹΑΔޮ཰తͳ৔֓೦ͷܗ
Efficient Spatial Concept Formation by Active Exploration of the Environment
୩ޱ জ ∗1
Akira Taniguchi
ాᔹ ٛج ∗1
Yoshiki Tabuchi
Τϧ ϋϑΟ ϩτϑΟ∗1
Lotfi El Hafi
ഡݪ ྑ৴ ∗1
Yoshinobu Hagiwara
୩ޱ ஧େ ∗1
Tadahiro Taniguchi
∗1໋ཱؗେֶ
Ritsumeikan University
Autonomous service robots are required to adaptively learn the categories and names of various places through the
exploration of the surrounding environment and interactions with users. In this study, we aim to realize the efficient
learning of spatial concepts by autonomous active exploration with a mobile robot. Therefore, we propose an active
learning algorithm that combines sequential Bayesian inference by a particle filter and position determination
based on information-gain in probabilistic generative models. Our experiment shows that the proposed method
can efficiently determine the position to form spatial concepts in simulated home environments.
1. ͸Ίʹ
αʔϏεϩϘοτΦϑΟε΍Ոఉ؀ڥʹΔʹ
ɼғͷ؀ڥ΍ͱͷΠϯλϥΫγϣϯΛ௨؀ڥʹଘࡏ
Δ༷ʑͳ৔ͷΧςΰϦ΍໊લΛࣗ཯తʹֶΔͱ
ཁͰΔɽϩϘοτ͸؀ڥΛҠಈதɼΒൃ࿩Ε৔ͷ
໊લ΍ɼͷҐஔͳͲͷ৔ʹؔΔϚϧνϞʔμϧσʔλ
ΛಘͰΔɽɼґવͱଟͷ৔߹Ͱ͸ɼֶ༻ͷ
σʔλΛΊΔΊʹϢʔβϩϘοτΛૢ࡞؀ڥ಺ΛҠ
ಈΔඞཁΔɽΕ͸ಛʹϩϘοτͷૢ࡞ʹ׳Εͳ
Ϣʔβ΁ͷෛ୲େͱߟΒΕΔɽͷΑͳ໰୊Λղ
ܾΔΊʹɼϩϘοτ͸ೳಈతʹ୳ࡧҐஔΛܾఆಈ࡞Δ
ͱٻΊΒΕΔɽ
ϩϘοτͷࣗ཯ΛߟΔͱಈతͳֶΑΓ΋ೳಈతͳ
ҙࢥܾఆʹجֶඞཁෆՄܽͰΔͱߟΔɽͰɼ
ಈతͱ͸ϢʔβϩϘοτΛಈ࡞Δͱɼೳಈతͱ͸
ϩϘοτࣗΒಈ࡞ΔͱͰΔɽΕ·Ͱզʑ͸ɼҐஔɾ
Իݴޠɾը૾ͷϚϧνϞʔμϧ৘ใʹجϊϯύϥϝτ
ϦοΫϕΠζ৔֓೦֫ಘϞσϧʹࣗݾҐஔఆͱ஍
(SLAM) [Thrun 05] Λ౷߹ SpCoSLAM [Taniguchi 17]
ΛఏҊɽ·ɼύʔςΟΫϧϑΟϧλʹجΦϯϥΠϯ
ڭࢣͳֶΞϧΰϦζϜʹΑΓະ஌؀ڥԼΒͷ஍ɾ৔
֓೦ɾޠኮͷஞֶ࣍ΛՄೳʹɽͰɼϩϘοτ؍ଌ
ϚϧνϞʔμϧ৘ใΒڭࢣͳֶʹΑΓܗΕ৔
ʹؔΔΧςΰϦ஌ࣝΛ৔֓೦ͱݺͿɽɼֶͷࡍ
͸ϢʔβϩϘοτΛ௥/ૢ࡞ɼڭࣔର৅ͷ৔΂ʹ
ରҠಈͱෳճͷൃ࿩ΛߦɽͷΑͳɼϩϘο
τʹͱಈతͳֶͰ͸Ϣʔβʹෛ୲ΛڧΔɼೳಈత
ͳֶͰ͸Ϣʔβ͸ϩϘοτʹ࣭໰Εͱʹൃ࿩ΔͷΈ
ͰΑɼϢʔβͷෛ୲ΛݮΒͱͰΔɽ
·ɼೳಈֶΛߦࡍ͸ɼֶ҆ఆతʹ౓ΑߦΔ
ংͰσʔλબ୒ΔͱཁͰΔɽύʔςΟΫϧϑΟϧ
λͷΑͳΦϯϥΠϯڭࢣͳֶͰ͸ɼ؍ଌσʔλͷ൪
ఆ౓ʹӨڹΛٴ΅ͱ஌ΒΕΔ [Ulker 10]ɽ
ͷΊຊݚڀͰ͸ɼҠಈϩϘοτʹΑΔࣗ཯తͳೳಈ୳ࡧ
ʹΑΓ৔֓೦ͷֶͷޮ཰ԽΛ࣮ݱΔͱΛ໨ࢦɽ
Ͱɼզʑ͸৔֓೦ܗͷΊͷ֬཰తϞσϧʹɼ
࿈བྷઌ:୩ޱজɼ໋ཱؗେֶ ৘ใཧ޻ֶ෦ɼ
a.taniguchi@em.ci.ritsumei.ac.jp
Select the position that maximize IG
Learn spatial concepts
Word
information
“There is
living room”
,  = (1.8, −2.0)
1: ϩϘοτʹΑΔೳಈతͳ୳ࡧͱֶͷ֓ཁ
ύʔςΟΫϧϑΟϧλʹΑΔஞ࣍ϕΠζ࿦ͱ৘ใརಘʹج
୳ࡧҐஔܾఆΛ૊Έ߹Θೳಈֶ๏ Spatial Con-
cept Formation with Information Gain-based Active Explo-
ration (SpCoAE) ΛఏҊΔɽຊݚڀͷ֓ཁΛ 1ʹࣔɽ
2. ؔ࿈ݚڀ
ೳಈ୳ࡧ΍ೳಈֶͷؔ࿈ݚڀͱɼSLAM ͷΊͷೳಈ
తͳ୳ࡧʹجҠಈઌҐஔܾఆ (Active SLAM) [Stachniss
05] ΍ɼϚϧνϞʔμϧΧςΰϦೝࣝͷΊͷೳಈ஌֮ (Active
Perception) [Taniguchi 18] Δɽ
[Stachniss 05] Ͱ͸ɼ ύʔςΟΫϧϑΟϧλʹجΦϯ
ϥΠϯ SLAM ͷҰͰΔ FastSLAM [Grisetti 05] ͷఆࣜ
Խͷ্Ͱෆ࣮֬ΛݮগΔΑͳҠಈઌΛܾఆΔɽҠಈ
ઌީิͷ͸๲େͱͳΔΊɼ୳ࡧީิ఺ΛߜΓࠐΉΊʹϑ
ϩϯςΟΞΞϓϩʔν [Yamauchi 98] ༻ΒΕΔɽ
[Taniguchi 18] Ͱ͸ɼϚϧνϞʔμϧΧςΰϦܗͷ֊૚ϕ
ΠζϞσϧͰΔ Multimodal Hierarchical Dirichlet Process
(MHDP) [Nakamura 11] ʹجɼ৮֮ɼࢹ֮ɼௌ֮ͱ
ײ֮ϞμϦςΟʹରԠΔߦಈબ୒Λߦɽ·ɼMHDP
1
The 34th Annual Conference of the Japanese Society for Artificial Intelligence, 2020
2M4-OS-3a-05

Ңಊ઎݀ఈ
Ңಊ
ϜϩοϠʖξϩ
৚ๅ͹
৖֕೨͹
Ψϱϧ΢ϱָ
ϫϚρφͶΓΖ
೵ಊద͵୵ࡩ
ϫϚρφ͹Ғ஖
ʶ
ϤʖδͶΓΖ
໌઴͹൅࿫
2: ೳಈతͳ୳ࡧʹΑΔ৔֓೦ܗͷྲྀΕ
ͰѻϞμϦςΟͷ͸ɼϩϘοτ؍ଌʹ༻Δηϯαͷ
ྨʹԠ֦ுΔͱͰΔɽ
ɼ[Stachniss 05] Ͱ͸৔ͷ໊લͷֶ΍ΧςΰϦܗ
͸ߦΘΕΒɼ[Taniguchi 18] Ͱ͸ط஌ͷ෺ମΧςΰ
ϦͷೝࣝͷΊͷߦಈબ୒ͰΓɼະ஌ͳ৔ΧςΰϦͷೳ
ಈֶ͸ߦΘΕͳɽ· [Taniguchi 18] ͸Gibbs
Sampling ʹجόονֶΞϧΰϦζϜΛલఏͱɽ
ೳಈֶʹ͸ɼຖճ৽ͳσʔλಘΒΕΔΊɼΦϯ
ϥΠϯֶͷํదͰΔͱߟΔɽ
ͷΑͳೳಈతͳߦಈબ୒ͷ๏ͷଟͰ͸ɼ৘ใརಘ࠷
େԽج༻ΒΕΔɽ৘ใརಘ࠷େԽج͸ྼϞδϡϥ
Λຬͱ஌ΒΕΓɼ৘ใརಘ࠷େͱͳΔҐஔͰ
ϢʔβΒ৔ͷ໊લΛಘΔͱͰɼ৔֓೦ܗʹΔෆ
࣮֬Λޮ཰తʹݮΒͱՄೳʹͳΔͱظ଴ΕΔɽ
[ాᔹ 18] Ͱ͸ɼϩϘοτ͸ Active SLAM ʹΑΓࣗ཯Ҡಈ
ΛߦͳΒ৔֓೦ͷܗΛɽϢʔβ͸ϩϘοτ
ಛఆͷ৔ʹདྷͱʹൃ࿩ڭࣔΛߦͱͰɼڭࣔʹରΔ
࣌ؒతෛ୲ͷܰݮΛ࣮ݱɽɼΕ͸஍ͷΊ
ͷೳಈ୳ࡧͰΓɼ৔֓೦ܗͷΊͷೳಈ୳ࡧ͸ߦΘΕ
ͳɽͷΊຊݚڀͰ͸ɼ৔֓೦ܗͷΊͷೳಈ
୳ࡧΛ࠷దԽ໰୊ʹؼணΔͱͰϩϘοτͷࣗ཯తͳ࣭໰
ߦಈͷҙࢥܾఆΛ࣮ݱΔɽ۩ମతʹ͸ɼ֊૚ϕΠζϞσϧͰ
Δ SpCoSLAM Λ୅දͱΔ৔֓೦ܗͷ֬཰తϞ
σϧʹɼ[Taniguchi 18] ͷೳಈ஌֮๏ͱಉ༷ͳఆࣜԽ
Λ৔֓೦ܗͷΊͷೳಈ୳ࡧʹରద༻Δɽ
3. ೳಈతͳ୳ࡧʹΑΔ৔֓೦ܗ
3.1 ֓ཁ
ຊߘͰ͸ɼ৔֓೦ܗͷ֬཰తϞσϧʹɼύʔ
ςΟΫϧϑΟϧλΑΔΦϯϥΠϯֶͱ৘ใརಘ (Information
Gain; IG) ʹجೳಈతͳ୳ࡧΛ૊Έ߹Θೳಈֶ๏
SpCoAE ΛఏҊΔɽఏҊ๏ʹΑΔೳಈֶͷྲྀΕΛ 2
ʹࣔɽ·ɼϩϘοτ؀ڥதͷҠಈઌީิΒ୳ࡧΔҠ
ಈઌΛܾఆΔɽҠಈޙɼϩϘοτ͸ “͸ͲͰʁ”
ͷΑʹʹ৔ͷ໊લΛͶΔͱͰɼͷҐஔͱ໊લΛಘ
Δɽ࣍ʹɼಘ৘ใΛݩʹ৔֓೦ܗΛߦɽɼܗ
Ε৔֓೦ʹج࣍ͷҠಈઌΛܾఆΔɽҎ্ͷྲྀΕΛ
ҠಈઌީิແͳΔ·Ͱ܁Γฦɽ
3.2 ֬཰తϞσϧ
3ʹ৔֓೦ܗͷάϥϑΟΧϧϞσϧΛɼද 1ʹ֤ม
ͷఆٛΛࣔɽΕ͸ SpCoSLAM [Taniguchi 17] Β SLAM
ͱޠኮ֫ಘΑͼը૾ಛ௃ʹؔΘΔมΛলུ୯ԽϞ
σϧͰΔɽάϥϑΟΧϧϞσϧͷփ৭ͷϊʔυ͸؍ଌมɼ
ന৭ͷϊʔυ͸ະ؍ଌมΛදɽN͸࠷తͳσʔλͷ؍
n
x
n
C
n
i
D
l
I
J
E
S
n
S
ܮ
l
W
00
,
N
m
00
,
Q
V
k
P
k
6
ܭ
ܰ
3: ৔֓೦ܗͷάϥϑΟΧϧϞσϧ
ද1: ৔֓೦ܗͷάϥϑΟΧϧϞσϧʹΔ֤ม
π৔֓೦ͷ index ͷଟ߲෼෍ύϥϝʔλ
Wl৔ͷ໊લͷଟ߲෼෍ύϥϝʔλ
φlҐஔ෼෍ͷ index ͷଟ߲෼෍ύϥϝʔλ
Cn৔֓೦ͷ index
Sn୯ޠ (৔ͷ໊લ)
inҐஔ෼෍ͷ index
μk,ΣkҐஔ෼෍ (ฏۉϕΫτϧɼڞ෼ࢄߦྻ)
xnϩϘοτͷҐஔ࠲ඪ
α, β, γ , ϋΠύʔύϥϝʔλ
m0,κ
0,V
0,ν
0
ଌɼLͱK͸ΕΕ৔֓೦ͱҐஔ෼෍ͷΧςΰϦΛ
දɽ·ɼϞσϧΛࣜ (1)-(9) ʹࣔɽ
π∼Dir(α)(1)
Wl∼Dir(β)(2)
φl∼Dir(γ)(3)
Cn∼Mult(π)(4)
Sn∼Mult(WCn)(5)
in∼Mult(φCn)(6)
Σk∼IW(V0,ν
0)(7)
μk∼N(m0,Σk/κ0)(8)
xn∼N(μin,Σin)(9)
Dir() ͸σΟϦΫϨ෼෍ɼMult() ͸ଟ߲෼෍ɼIW() ͸ٯ
΢Ογϟʔτ෼෍ɼN() ͸Ψ΢ε෼෍ͰΔɽ
3.3 ΦϯϥΠϯֶΞϧΰϦζϜ
ຊݚڀͰ͸ɼϩϘοτ࣍ͷҠಈઌΛબ୒ΔΊʹɼ؍ଌ
৘ใΒ৔֓೦ΛֶΔɽͷΊɼSpCoSLAM ͱಉ༷
ʹɼRao-Blackwellized Particle Filter Λ༻ΦϯϥΠϯֶ
Λߦɽຊ๏Ͱ͸ɼ৔ʹؔΔϚϧνϞʔμϧ৘ใͱ
৔ͷޠኮ৘ใͱҐஔ৘ใΛ؍ଌͱɼ৔֓೦ͷ֤ύϥ
ϝʔλͷࣄޙ෼෍ΛఆΔɽ
৔֓೦ΛֶΔࡍʹఆΔ΂ͷύϥϝʔλͷಉ
࣌ࣄޙ෼෍ΑͼͷҼࢠ෼ղ͸ҎԼͷΑʹͳΔɽ
p(Θ,C
1:n,i
1:n|x1:n,S
1:n,h)
=p(Θ |C1:n,i
1:n,x
1:n,S
1:n,h)
p(C1:n,i
1:n|x1:n,S
1:n,h) (10)
ͳɼ৔֓೦ͷ֤ύϥϝʔλ߹Λ Θ={μ, Σ,φ,W,π}ɼϋ
Πύʔύϥϝʔλ߹Λ h={α, β, γ , m0,κ
0,V
0,ν
0}ͱɽ
2
The 34th Annual Conference of the Japanese Society for Artificial Intelligence, 2020
2M4-OS-3a-05

3.4 ೳಈ୳ࡧΞϧΰϦζϜ
৘ใཧ࿦త౓ͷҰͰΔ৘ใརಘ IG Λ༻؀ڥ಺Ͱ
࠷΋ෆ࣮֬ΛݮগͰΔͰΖҐஔΛબ୒ɾҠಈɼڭࣔ
ൃ࿩ΛؚΉ؍ଌ৘ใΛಘΔɽ
ఏҊ๏Ͱ͸ɼσʔλΛ Nճ؍ଌͱͷ࠷తͳࣄ
ޙ෼෍ͱɼݱࡏͷεςοϓΒ࣍ͷҠಈઌͰσʔλΛ؍ଌ
ͱͷࣄޙ෼෍ͷؒͷ Kullback–Leibler divergence (KL
divergence) Λ࠷খԽΔσʔλ aΛબ୒ΔఆࣜԽΛͱΔɽ
ͳɼn0⊂1:N͸ 3ʹΔطʹ؍ଌΕϓϨʔτ൪
߸ͷ߹ͰΓɼະ؍ଌͷ৔߹͸ n0=∅ͱΔɽ
minimize
aDKL(p(Θ,C
1:N,i
1:N|x1:N,S
1:N,h)
p(Θ,C
1:N,i
1:N|xn0∪a,S
n0∪a,h)) (11)
ɼͷ x1:N,S
1:N΋xn0∪a,S
n0∪aͷΕ΋࣍ͷҠ
ಈઌʹҠಈΔલʹ؍ଌΔͱ͸ͰͳɽͰదͳ୅
ସҊͱࣜ (12) ͷΑʹ KL divergence ͷظ଴஋Λܭࢉ
ΔͱͰ࣍ͷҠಈઌͰ؍ଌΕΔσʔλͷ൪߸ a∗ΛٻΊΔɽ
ͷͱɼa∈1:N\n0͸ɼ؍ଌࡁΈσʔλΛআɼ࣍ͷҠಈ
ઌͰ؍ଌΕΔσʔλͷ൪߸ͰΔɽͷࣜΛมܗΔͱɼࣜ
(13) ͷΑʹɼ࣍ͷҠಈઌͰσʔλΛ؍ଌޙͷࣄޙ෼෍
ͱɼݱࡏͷεςοϓͰͷࣄޙ෼෍ͷؒͷ KL divergence ͷظ
଴஋Λ࠷େԽΔࣜͱͳΔɽɼͷ KL divergence ͷ
ظ଴஋Λ৘ใརಘ IG ͱΔͱɼࣜ (14) ͷΑʹදΔɽͳ
ɼZ={Θ,C
1:N,i
1:N}ɼXn0={xn0,S
n0,h}ͱɽ
a∗= argmin
a
EX1:N\n0|Xn0
[DKL(p(Z|X1:N)p(Z|Xn0∪a))] (12)
= argmax
a
EXa|Xn0[DKL(p(Z|Xn0∪a)p(Z|Xn0))]
(13)
= argmax
a
IG(Z;Xa|Xn0) (14)
্هͷಋ͸ Active Perception [Taniguchi 18] ʹΔ
ಋͱಉ༷ͰΔɽͰɼIG ͷ۩ମతͳܭࢉࣜΛҎԼʹࣔ
ɽ[Taniguchi 18] ʹϞϯςΧϧϩۙࣅΛߦՕ
͸ɼຊఏҊ๏Ͱ͸ࣜ (16) ͷΑʹɼύʔςΟΫϧϑΟϧ
λͷఆ݁ՌΛར༻ΔͱͰΔɽ·ɼXaʹؔ͸ɼ
ࣜ(17) ͷΑʹ༧ଌ෼෍Β Jݸͷٙࣅ؍ଌΛαϯϓϦϯά
ΔͱͰۙࣅΔɽͷͱɼR͸ύʔςΟΫϧͷݸͰ
Γɼω[r]
n0͸ύʔςΟΫϧͷΈͰΔɽ
IG(Z;Xa|Xn0)
=
Z
Xap(Z, Xa|Xn0)log p(Z, Xa|Xn0)
p(Z|Xn0)p(Xa|Xn0)(15)
≈
R

r=1 
Xap(Xa|Z[r],X
n0)ω[r]
n0
log p(Xa|Z[r],X
n0)
R
r=1 p(Xa|Z[r],X
n0)ω[r]
n0,Z
[r]∼q(Z|Xn0)
(16)
≈
R

r=1
J

j=1 ⎡
⎣ω[r]
n0log p(X[j]
a|Z[r],X
n0)
R
r=1 p(X[j]
a|Z[r],X
n0)ω[r]
n0⎤
⎦,
X[j]
a∼p(Xa|Z[r],X
n0) (17)
(a) ࢖༻Ոఉ؀ڥ (b) SLAM ʹΑΓ࡞஍
4: SIGVerse ্Ͱͷ࣮ݧ؀ڥ
4. ࣮ݧ
ఏҊ๏ʹΑΔೳಈతͳ୳ࡧʹΑޮ཰తͳ৔֓೦ܗ
࣮ݱͰΔʹධՁΔɽ
4.1 ࣮ݧ৚݅
ຊ࣮ݧͰ͸ɼ࣮ݧ؀ڥͱγϛϡϨʔλ SIGVerse [஍ 17]
্Ͱߏங 4(a) ͷՈఉ؀ڥʹɼࣄલʹ࡞ 4(b) ͷ
஍Λ༻Δɽ஍ͷ࡞ʹ͸ɼGrid-based FastSLAM 2.0
[Grisetti 05] ͷΞϧΰϦζϜͰಈ࡞Δ ROS ύοέʔδͷ
gmapping Λ࢖༻ɽϩϘοτͱɼSIGVerse ্Ͱಈ࡞
ΔHuman Support Robot (HSR) ͷԾ૝ϞσϧΛ࢖༻ɽ
σʔλΛ؍ଌΔҐஔ࠲ඪͷީิ͸ɼന৭ͰදΕΔ஍
্ͷϑϦʔεϖʔεʹ 0.8 m ִؒͰઃఆɽͷதͰɼ௚
ܘ0.43 m ͷ୆ΛඋΔ HSR ҠಈΔεϖʔεΛ֬อ
ΔΊʹɼҐஔ࠲ඪΒ 0.5 m Ҏ಺ʹน΍ো֐෺Δ৔
߹͸ͷҐஔ࠲ඪΛ࡟আɽύʔςΟΫϧ͸ R= 1000ɼ
ٙࣅ؍ଌ͸ J=10ͱɽϋΠύʔύϥϝʔλ͸ɼα=
1.0,β=0.01,γ=0.1,m
0=[0.0,0.0]T,κ
0=0.001,V
0=
diag(1.5,1.5),ν
0=4.0ͱɽ
ධՁࢦඪͱɼΫϥελϦϯάͷೳΛଌΔࢦඪͷҰͰ
Δ Adjusted Rand Index (ARI) Λ࠾༻ΔɽఏҊ๏͸Φ
ϯϥΠϯֶͰΔΊɼ؍ଌࡁΈσʔλͷΈʹରΔ ARI
(Step-by-step) ͱɼະ؍ଌͷσʔλʹର֤εςοϓʹ
ఆΕ৔֓೦ͷύϥϝʔλ Θ={μ, Σ,φ,W,π}ͷࣄ
ޙ෼෍Λݩʹ༧ଌ݀ຒΊ ARI (Predictive padding) ͷ
ೋྨͷධՁํ๏Λ༻Δɽ৔֓೦ͷ index CnͱҐஔ෼෍
ͷindex inͷΕΕʹධՁΔɽ
ൺֱ๏͸ɼҎԼͷ 3ͱΔɽ
(A) SpCoAEɿఏҊ๏
(B) RandomɿϥϯμϜʹҠಈΔҐஔΛબ୒Δ๏
(C) IG minɿIG খʹ୳ࡧҐஔΛબ୒Δ๏
4.2 ࣮ݧ݁Ռ
ఏҊ๏ʹΑΔஞ࣍తͳ৔֓೦ܗͷ༷ࢠͷҰྫΛ 5ʹ
ࣔɽ֤ପԁ͸ఆΕҐஔ෼෍ΛදɽՄࢹԽͷࡍɼ৭͸
ϥϯμϜʹܾఆΕɽ֤ପԁ಺ͷ఺͸ɼϩϘοτΕ·Ͱ
σʔλΛಘҐஔΛࣔΔɽStep 25 Β Step 28 Ͱ
͸ɼࠨԼͷ 3Λ࿈ଓ୳ࡧΔɽ·ɼStep 36 
ΒStep 40 Ͱ͸ɼΜத্෦ͷ෦԰Λ఺తʹ୳ࡧΔɽ
ͷͱΒɼఏҊ๏͸৔ͷ໊લڭࣔΕΒෆ֬
࣮ͳՕΛճ·ͱΊ୳ࡧΔͱʹΑΓɼͷ৔ʹؔ
Δෆ࣮֬Λ෼ʹݮগޙʹɼผͷ৔ͷ୳ࡧʹҠΔΑ
ͳڍಈΛࣔͱΘΔɽ
֤๏ʹରɼશ 10 ࢼߦͷ֤ධՁ஋ͷฏۉ஋ͱඪภࠩ
ͷΤϥʔόʔΛ 6ʹࣔɽ࣠ ARIɼԣ࣠εςοϓ
ͰΓɼઢ SpCoAEɼ྘ઢ Randomɼઢ IG min
ͰΔɽARI (Step-by-step) ʹؔ͸ɼશମతʹ SpCoAE
ଞͷ๏ͱൺֱߴ஋ΛࣔΔͱΘΔɽΕ
3
The 34th Annual Conference of the Japanese Society for Artificial Intelligence, 2020
2M4-OS-3a-05

Step 36 Step 40
Step 25 Step 28
5: ೳಈతͳ୳ࡧʹΑΔ৔֓೦ܗͷ༷ࢠ
(a) ৔֓೦ͷ index CʹରΔ
ARI (Step-by-step)
(b) Ґஔ෼෍ͷ index iʹରΔ ARI
(Step-by-step)
(c) ৔֓೦ͷ index CʹରΔ ARI
(Predictive padding)
(d) Ґஔ෼෍ͷ index iʹରΔ ARI
(Predictive padding)
6: εςοϓͱͷ ARI ͷҠ
͸ɼ֤εςοϓʹΑΓ֬ʹ৔֓೦ΛֶͰΔΑ
ͳҠಈઌΛબ୒ͰΔΒͱߟΔɽARI (Predictive
padding) ʹؔ͸ɼظεςοϓͰ͸ SpCoAE ͸Random
ΑΓׯྼΔ΋ͷͷɼޙ൒ͷεςοϓͰ͸࠷΋ߴ஋
ΛࣔɽΕ͸ɼ؀ڥશମΛૉૣ໢ཏΔ఺ʹ͸
Random ༗ޮͰΔɼΑΓ֬ͳ৔֓೦ΛֶΔ఺
ʹ͸ SpCoAE ༗ޮͰΔͱΛࣔΔɽ
5. ΘΓʹ
ຊߘͰ͸ɼϩϘοτͷ৔֓೦ܗʹؔΔೳಈతͳ୳ࡧ
๏ΛఏҊɽఏҊ๏͸৘ใརಘ IG ࠷େͱͳΔީิ఺Λ
બ୒ɼҠಈઌʹಘ৘ใΒΦϯϥΠϯֶΛߦ
ͱʹΑΓೳಈֶΛ࣮ݱɽ࣮ݧͰ͸ɼΑΓ֬ͳ৔֓೦
ͷܗΛߦΔͱҙຯʹɼೳಈతͳ୳ࡧͷ༗ޮΛ
ࣔɽࠓޙ͸ɼΑΓఏҊ๏ͷޮ཰ΛࣔΊʹɼҠಈڑ
཭Λߟྀൺֱ࣮ݧͳͲΛߦ༧ఆͰΔɽ
ຊߘͰ͸ɼSpCoSLAM ΒҰ෦ͷมΛলϞσϧʹΑ
ΔೳಈֶΛ࣮ݱɽࠓޙͷల๬ͱɼSpCoSLAM ͷ֬
཰తϞσϧ΁ͷೳಈֶͷద༻ڍΒΕΔɽͷΊ
ʹɼը૾ಛ௃ΛؚΉީิબ୒Αͼɼ஍Λ࣋ͳະ஌؀ڥ
ʹΔ Active SLAM ͱͷ౷߹Λߦɽ
ࣙ
ຊݚڀͷҰ෦͸ɼJST CREST JPMJCR15E3 ͷΛ
΋ͷͰΔɽ
ࢀߟจݙ
[Grisetti 05] Grisetti, G., Stachniss, C., and Burgard, W.:
Improving Grid-based SLAM with Rao-Blackwellized
Particle Filters by Adaptive Proposals and Selective Re-
sampling, in IEEE ICRA (2005)
[Nakamura 11] Nakamura, T., Nagai, T., and Iwahashi, N.:
Multimodal categorization by hierarchical Dirichlet pro-
cess, in IEEE/RSJ IROS, pp. 1520–1525 (2011)
[Stachniss 05] Stachniss, C., Grisetti, G., and Bur-
gard, W.: Information Gain-based Exploration Using
Rao-Blackwellized Particle Filters., in Robotics: Science
and Systems, Vol. 2, pp. 65–72 (2005)
[Taniguchi 17] Taniguchi, A., Hagiwara, Y., Taniguchi, T.,
and Inamura, T.: Online Spatial Concept and Lexical Ac-
quisition with Simultaneous Localization and Mapping,
in IEEE/RSJ IROS, pp. 811–818 (2017)
[Taniguchi 18] Taniguchi, T., Yoshino, R., and Takano, T.:
Multimodal Hierarchical Dirichlet Process-Based Ac-
tive Perception by a Robot, Frontiers in Neurorobotics,
Vol. 12, p. 22 (2018)
[Thrun 05] Thrun, S., Burgard, W., and Fox, D.: Proba-
bilistic Robotics, MIT Press (2005)
[Ulker 10] Ulker, Y., G¨unsel, B., and Cemgil, T.: Sequen-
tial Monte Carlo samplers for Dirichlet process mixtures,
in AISTATS, pp. 876–883 (2010)
[Yamauchi 98] Yamauchi, B.: Frontier-based Exploration
Using Multiple Robots, in AGENTS, pp. 47–53 (1998)
[஍ 17] ஍ ྑ໌,Ҵ༢ ఩໵ɿUnity ͱROS Λ౷߹Ϋ
ϥ΢υܕϚϧνϞʔμϧର࿩ܦݧ஝ϓϥοτϑΥʔϜ,
޻஌ೳֶձશࠃେձ࿦จ, No. 4G1OS14a4 (2017)
[ాᔹ 18] ాᔹٛج,୩ޱজ,ഡݪྑ৴,୩ޱ஧େɿՈఉ؀ڥʹ
ΔҠಈϩϘοτͷೳಈత஍ͱ৔֓೦ܗ,޻
஌ೳֶձશࠃେձ࿦จ, No. 2L3OS6b03 (2018)
4
The 34th Annual Conference of the Japanese Society for Artificial Intelligence, 2020
2M4-OS-3a-05