Long-Term Autonomy in Everyday Environments - bcs.org · Haus der Barmherzigkeit, Vienna, Austria...
Transcript of Long-Term Autonomy in Everyday Environments - bcs.org · Haus der Barmherzigkeit, Vienna, Austria...
Nick Hawes
Long-Term Autonomy in Everyday Environments
School of Computer Science, University of Birmingham, UK
A New Challenge for AI and Robotics
http://nickhaw.es @hawesie
Long-Term Autonomy in Everyday
Environments
http://strands-project.eu
Robust, intelligent,
autonomous behaviour
Long run-times in everyday
environments
Novel opportunities
to learn structure
environment
Exploitation of structure for
improved performance
A New Challenge for AI and Robotics
Long run-times in everyday
environments
Exploitation of structure for
improved performance
A New Challenge for AI and Robotics
Meta-room mapping Desktop observations
Object presence checks Door checks
G4S Technology, UK
Haus der Barmherzigkeit,
Austria
Information provision Object presence checks
Door checks
G4S Technology, Challenge House, Tewkesbury, UK 690m3
Haus der Barmherzigkeit, Vienna, Austria 1030m3
Task ActionTask Action
Continuous
Topological
Monitoring
Nav Learning
Routine
Task Executor
Task Action
Scheduler
Task Action
Localisation& Navigation
ExecutiveControl
ApplicationSpecific
Task ActionTask Action
Continuous
Topological
Monitoring
Nav Learning
Routine
Task Executor
Task Action
Scheduler
Task Action
Localisation& Navigation
ExecutiveControl
ApplicationSpecific
Continuous
Continuous
Topological
Continuous
Topological
Monitoring
Continuous
Topological
Monitoring
Nav Learning
Continuous
Topological
Monitoring
Nav Learning
Task Executor
Continuous
Topological
Monitoring
Nav Learning
Task Executor
Routine
Check fire doors Check fire extinguisher
Check all doors Observe desks Patrol corridors
Check fire doors Map offices
Check all doors Observe desks Patrol corridors
Charge
Upload data Replicate database
Process maps
From 9:00 to 17:00 Weekdays, except 26/5/14
Continuous
Topological
Monitoring
Nav Learning
Task Executor
Routine
Schedulertask
task
task
task
task
Localisation& Navigation
ExecutiveControl
ApplicationSpecific
Task ActionTask Action
Continuous
Topological
Monitoring
Nav Learning
Routine
Task Executor
Task Action
Scheduler
Task Action
Care Security
Deployment 14/5/14 to 4/6/14 22/5/14 to 12/6/14
Working Hours Weekdays, 8.00 to 17.00 Weekdays, 8.45am to 17.45
Distance 27.94km 20.64km
Tasks Completed 1985 963
Autonomous Time 48h 53m 17s 26h 18m 51s
System Lifetime
Max SL 171h 0m (7d 3h 0m) 91h 0m (3d 19h 0m)
Max SL working 48h 40m (2d 0h 40m) 39h 30m (1d 15h 30m)
wait object check door check metric map desktop perception
wait patrol object check idle/engagement door check
G4S Technology, UK
Haus der Barmherzigkeit,
Austria
Long run-times in everyday
environments
Exploitation of structure for
improved performance
A New Challenge for AI and Robotics
mean time from robot
straight line time
Best 8 matches between straight-line and recorded times
mean time from robot
straight line time
Worst 8 matches between straight-line and recorded times
W1 W2
W3
0.9
action goto W2 from W1
0.1
cost mean time from all attempts
W1 W2
W3
0.9
e.g. (F W2) (eventually reach W2)
0.1
express navigation goals in Linear Temporal Logic
W1 W2
W3
0.9
0.1
W2¬W2 true
╳
B. Lacerda, D. Parker, and N. Hawes. Optimal and Dynamic Planning for Markov Decision Processes with Co-Safe LTL Specifications. In: IROS 2014.
B. Lacerda, D. Parker, and N. Hawes. Optimal and Dynamic Planning for Markov Decision Processes with Co-Safe LTL Specifications. In: IROS 2014.
Qualitative Spatial Relations (QSRs)
Akshaya Thippur et al. KTH-3D-TOTAL: A 3D Dataset for Discovering Spatial Structures for Long-Term Autonomous Learning. In SAIS’14.
Lars et al. Bootstrapping probabilistic models of qualitative spatial relations for active visual object search. In AAAI SS 2014 on Qualitative Representations for Robots
Object Presence Probability Pr
obab
ility
0
0.25
0.5
0.75
1
Mon
itor
Keyb
oard
Mou
se
Telep
hone
Cup/
Mug
Pen/
Penc
il
Book
Bottl
e
Desk
top
PC
Stap
ler
Lapt
op
Mob
ile p
hone
Lam
p
Calcu
lator
Keys
Head
phon
e
Glas
s
Hole
punc
h
0.0
0.5
1.0
left right front behind close distant
0.0
0.5
1.0
left right front behind close distant
0.0
0.5
1.0
left right front behind close distant
0.0
0.5
1.0
left right front behind close distant
0.0
0.5
1.0
left right front behind close distant
book wrt. monitor
mug wrt. monitor
PC wrt. monitor
keyboard wrt. monitor
mouse wrt. monitor
Position of cup relative to monitor
Position of cup relative to keyboard
Supporting planes vs QSRs 10 trials 3 out of 8 tables choose 1/500 sim. desks
L. Kunze, K. K. Doreswamy and N. Hawes. Using Qualitative Spatial Relations for Indirect Object Search. In ICRA’14.
0.0
17.5
35.0
52.5
70.0
0
2.5
5
7.5
10
Random Views Supporting Planes Correct QSRs Partially Correct QSRs Misleading QSRs
Objects Found (/10) Time (secs) Poses
3.23.1
1.1
2.3
4.8
65.0
55.0
15.6
33.6
68.5
6
8
1010
6
Search Results (Simulation)
Search Results (Robot)
0.0
17.5
35.0
52.5
70.0
0
2.5
5
7.5
10
Supporting Planes Correct QSRs
Objects Found (/10) Time (secs) Poses
1.1
2.2
33.4
69.5 10
9
Qualitative Spatial Relations (QSRs)
train: 19 desks, 3 scenes per desk = 57 scenes test: 1 desk, 3 scenes per desk = 3 scenes
Lars Kunze et al. Combining Top-down Spatial Reasoning and Bottom-up Object Class Recognition for Scene Understanding. In IROS ’14.
0.0
25.0
50.0
75.0
100.0
No Relations Learnt Metric Relations Ternary Point Calculus Ternary Point Calculus Ternary Point Calculus Distance
Relative Size
Ternary Point Calculus Distance
Relative Size Connectivity
88.9490.98
54.72
45.38
95.65
0
89.992.3
65.059.2
96.0
59.2
With Visual Classification Without Visual Classification
Classification Results (Robot)
Rares Ambrus et al. Meta-Rooms: Building and Maintaining Long Term Spatial Models in a Dynamic World. In IROS ’14.
Rares Ambrus et al. Meta-Rooms: Building and Maintaining Long Term Spatial Models in a Dynamic World. In IROS ’14.
Rares Ambrus et al. Meta-Rooms: Building and Maintaining Long Term Spatial Models in a Dynamic World. In IROS ’14.
Rares Ambrus et al. Meta-Rooms: Building and Maintaining Long Term Spatial Models in a Dynamic World. In IROS ’14.
Long run-times in everyday
environments
Exploitation of structure for
improved performance
A New Challenge for AI and Robotics
A New Challenge for AI and Robotics
Robust, intelligent,
autonomous behaviour
Long run-times in everyday
environments
Novel opportunities
to learn structure
environment
Exploitation of structure for
improved performance
A New Challenge for AI and Robotics
http://strands-project.eu
Nick Hawes
Long-Term Autonomy in Everyday Environments
School of Computer Science, University of Birmingham, UK
A New Challenge for AI and Robotics
http://nickhaw.es @hawesie