7/7/2005 1

Subterranean RoboticsFRC Seminar Series

Towards Topological Exploration of Abandoned Mines

Aaron Morrisacmorr@ri.cmu.edu

7/7/2005 2

Subterranean RoboticsAgenda

• Motivation• Key challenges• Existing solutions• Topologies, intersections, and navigation• Experimental Results• Conclusions• Things to come…

7/7/2005 3

Subterranean RoboticsThe Silent Subterranean Menace

Caves SewersMines

7/7/2005 4

Subterranean RoboticsTrouble Below The Surface

Subsidence

Encroachment

Rescue

Collapse

ImpoundmentMine Fire

7/7/2005 5

Subterranean RoboticsMitigation Through Information

• Poor condition• Idealized• Misaligned• Partially complete• Unavailable

Unfortunately…

Subterranean maps are the primary source of information on cavity extent.

7/7/2005 6

Subterranean RoboticsCurrent Data Acquisition Methods

• Places inspector into harms way

• Expensive• Not always an

available option

• Limited range• Time consuming• Void must be

inferred• Accuracy?

Direct Observation

• Requires many holes– Expensive– Time consuming

• Little quantitative information

Remote SensingRemote ObservationRemote Observation Remote Sensing

7/7/2005 7

Subterranean RoboticsRobotic Solutions

Why Robots?• Physical presence

without human risk• High fidelity data• Operation in the

harshest of conditions• Log and recall all

sensory input

7/7/2005 8

Subterranean RoboticsKey Challenges

Reliable localizationHighly cyclic

Physical constraints for portal entry, borehole, puddle, …Accessibility

Recognize and avoid obstacles beyond ground planeComplex 3D obstacles

Recognize and avoid obstacles, surmount everything elseRugged terrain

Reliable failure recovery, adequate sensing, …No communication

Roll, walk, swim, climb, …Harsh environment

Robotic ChallengeCondition

7/7/2005 9

Subterranean RoboticsRobotic Systems

Ferret IIGroundhog

Cave Crawler

Helix

Mine FishFerret III

7/7/2005 10

Subterranean RoboticsMaps and Models

Mathies Experiments [Thrun, Hänel, Montemerlo]

Bruceton Experiments Kansas City Void Bruceton Model

7/7/2005 11

Subterranean RoboticsPortal Inspection System

7/7/2005 12

Subterranean RoboticsThe Challenges Addressed

• Mobile platform• Portal Entry• Robust to certain

failures• Adequate sensor

configuration• Reliable autonomy• Compelling maps

7/7/2005 13

Subterranean RoboticsAccomplishments

Summary of Field Deployments in Mathies Mine

7/7/2005 14

Subterranean RoboticsThe Next Step: Network Exploration

Exploration and localization through topological representations

7/7/2005 15

Subterranean RoboticsPhase 1: Intersection Detection

P points

DT - DelaunayTriangulation

T triangles

DT(P)→T:∃tijk has corners

pi, pj, pk and edgeseij, eik, ejk

Intersections correspond to nodes of the GVD [Choset]

7/7/2005 16

Subterranean RoboticsOffline ID Results

100 Meters of Mine Corridor

In total, 9 features: 5 intersections and 4 pockets of excavated coal

7/7/2005 17

Subterranean RoboticsNode Classification

Not an intersection Could be? An intersection

7/7/2005 18

Subterranean RoboticsIntersection Detection Algorithm

Strong Node

• Features selected from DT of range scan, filtered by edge length

• Tracked over multiple scans• Groundhog sensor

configuration requires it to drive through intersection

• Identified as “strong” or “weak” (known as RGVD)

Weak Node

7/7/2005 19

Subterranean RoboticsPhase 2: Intersection Navigation

Elements of Navigation

• Workspace Computation

• Goal Selection

• Path Generation

No two intersections are alike. Path plans are likely not to be identical.

7/7/2005 20

Subterranean RoboticsWorkspace Computation

3D point set Local Gradient

for each p in 3D_Point_Cloudmap_cell_max[p.x, p.y] = MAX(p.z, map_cell_max[p.x, p.y])map_cell_min[p.x, p.y] = MIN(p.z, map_cell_min[p.x, p.y])

end forfor each r in Number_Map_Cell_Rows

for each c in Number_Map_Cell_Columnsif map_cell_min[r,c] > Max_Traverse_Height or

map_cell_max[r,c] < Platform_Height + Safety_Margin or(map_cell_min[r,c] or map_cell_max[r,c]) == No_Value

map_cell[r,c] = lethalelse

map_cell[r,c] = map_cell_max[r,c] - map_cell_min[r,c]

Cost map

7/7/2005 21

Subterranean RoboticsGoal Selection

Where do we go next?

Cost map Free space

7/7/2005 22

Subterranean RoboticsGoal Selection

X

X

uiuj

uk

nijk

g´j

g´k

gj

gk

tijkGiven:Voronoi node – nijkDelaunay Triangle – tijkPotential Goals – P

Calculate:unit vectors – ui uj ukgoal positions – g´j g´k

g´j = a·uj+ nijkg´k = a·uk+ nijk

Choose:gj gk such that

gj = minn=1..N(D(g´j ,Pn))gk = minn=1..N(D(g´k ,Pn))

7/7/2005 23

Subterranean RoboticsNavigation at an Intersection

3D Configuration Space

Nonholonomic Motions

7/7/2005 24

Subterranean RoboticsNonholonomic Motion Planning - GS

Robot Motion Planning, Jean-Claude Latombe, Kluwer, 1991.

θcosvx =& θsinvy =& φθ tanLv

=&

},0,{},{ maxmax00 φφ +−×− vv

Velocity Parameters

Control Parameters

Arc Template

0 where},,...,0,...,,{},{

..1max

max1max00

≠>

++−−×−

N

Nvvφφ

φφφφ

7/7/2005 25

Subterranean RoboticsC-Space

Collision Checking3 Dimensional Configuration SpaceC

R

W

R×C×W where W is [-π, π)and cell size is 10cm × 10cm

7/7/2005 26

Subterranean RoboticsThe Algorithm

qi

Priority Queueqi+1

qi

qi+2

qi+1

cj

qi+1

Priority Queue

f(c)=g(c)+h(c)

C-space

Open

Closed

7/7/2005 27

Subterranean RoboticsCost Metric

f(c)= g(c) + h(c)h(c) – heuristic value, D({xc,yc},G)

g(c) – cost of path from S to cg(c) = g(p) + g(p,c)g(p,c) = arc_length

7/7/2005 28

Subterranean RoboticsAlgorithm Characteristics

• Best-first search– “Obvious” solutions found quickly– More complicated maneuvers

take longer• Goal threshold

– Tighter thresholds longer search times and “quirky” paths

– Looser thresholds produce worse end poses

• Search time proportional with amount of free space

7/7/2005 29

Subterranean RoboticsExample path

Actual traversalSimulated traversal

7/7/2005 30

Subterranean RoboticsPhase 3: Topological Planner

7/7/2005 31

Subterranean RoboticsResults

Longest Autonomous Traverse to Date

• 2 hours to complete• Over 400 m• 8 nodes (3 intersections)

7/7/2005 32

Subterranean RoboticsOther Interesting Results

• Over 20 hours of operation, identified 50 intersections

• Strong node identification to total intersections: 100%

• Weak nodes identified as strong: 0• Average time to calculate motion plan: 10s (vs.

60s with former planner)• 30s to 60s for complex turning maneuvers• Complete autonomous tree exploration

7/7/2005 33

Subterranean RoboticsIn the Future…

• Intersection classification (spin images)• Probabilistic T-SLAM• Fault detection and state tracking• Framework for degraded operation modes• Implementation on new systems• Response and rescue• International endeavors

7/7/2005 34

Subterranean Robotics16-899C Subterranean Robotics

Mine Fire Response and Rescue

NSH 1109, MW 1:30-2:50PM

7/7/2005 35

Subterranean RoboticsContact Information

Subterranean Robotics Onlinewww.minemapping.org

www.subterraneanrobotics.org

Aaron Morris: acmorr@ri.cmu.eduScott Thayer: sthayer@ri.cmu.edu