Download presentation
Presentation is loading. Please wait.
Published byKevin Gaines Modified over 7 years ago
1
Trey Smith, Robotics Institute, Carnegie Mellon University
Rover Science Autonomy Probabilistic Planning for Science-Aware Exploration Thesis Proposal Trey Smith, Robotics Institute, Carnegie Mellon University 14 June 2004
2
Robotic Planetary Science
3
Mars Rover Operations Today
4
Mars Rover Operations Today
5
Mars Rover Operations Today
6
Over-the-Horizon Science
60 m
7
Over-the-Horizon Science
60 m 6 km
8
Science Autonomy The opportunity: Long-distance mobility gives the rover access to new sites The challenge: Do useful science in over-the- horizon situations The method: Enable the rover to reason about science goals and the science data it collects
9
New Operational Modes !
10
Sensing Uncertainty type 4 type 7 novel
11
Probabilistic Sensor Planning
contact sensing spectrometer keep moving keep moving
12
Probabilistic Sensor Planning
contact sensing spectrometer keep moving keep moving
13
POMDP Model Principled method for reasoning about information gain and future beliefs
14
POMDP Model Principled method for reasoning about information gain and future beliefs Horribly, horribly intractable
15
POMDP Model Principled method for reasoning about information gain and future beliefs Horribly, horribly intractable Develop approximation techniques that take advantage of structure in the SA domain to speed up planning
16
Thesis Statement Enabling a robot to reason about science goals and the science data it collects will significantly improve its exploration efficiency,
17
Thesis Statement Enabling a robot to reason about science goals and the science data it collects will significantly improve its exploration efficiency, and POMDP methods can significantly improve the quality of science autonomy plans.
18
Outline Motivation Background and related work Technical approach
Preliminary work Proposed research
19
Science Autonomy Related Work
Image and spectral analysis [Gulick et al., 2001] [Gazis and Roush, 2001] Balancing exploration priorities [Moorehead, 2001] Sensor planning to detect meteorites [Wagner et al., 2001] Orbiter reacting to science events [Chien et al., 2003] OASIS project: closed-loop SA system [Estlin et al., 2003]
20
POMDP Introduction Recall: Great for sensor planning
Tend to be massively intractable
21
Example Problem: RockSample
22
Example Problem: RockSample
23
Example Problem: RockSample
exit
24
Example Problem: RockSample
exit check
25
Example Problem: RockSample
exit check good bad (+20 or +0) (+10) = 0.9
26
History and Belief initially 50/50 check good check good 0.5 0.7 0.9
bad good Pr(good) = 0.5 Pr(good) = 0.7 Pr(good) = 0.9 check good check good
27
Policy exit check sample sample check bad good good 0.38 0.64 0.5 bad
0.7 bad
28
Value Function: Horizon 1
exit sample 20 10 check bad good
29
Value Function: Horizon 1
exit sample 20 10 bad good
30
Value Function: Horizon 2
exit sample 20 check bad good 10 exit sample bad good
31
Computing the Policy exit check sample 20 10 exit sample check bad
good 10 exit sample exit check sample
32
optimal value function V*
Value Iteration Horizon 2 optimal value function V* Horizon 1 Horizon 0
33
POMDP Related Work Policy iteration [Hansen, 1998] [Poupart and Boutilier, 2003] Gradient ascent policy search [Baxter and Bartlett, 2000] Value iteration [Sondik, 1971]
34
Value Iteration Variants
Belief Space
35
Value Iteration Variants
Exact value iteration [Littman, 1994] [Cassandra et al., 1997]
36
Value Iteration Variants
Exact value iteration [Littman, 1994] [Cassandra et al., 1997]
37
Value Iteration Variants
38
Value Iteration Variants
Grid-based [Brafman, 1997] [Hauskrecht, 2000]
39
Value Iteration Variants
What next?
40
Value Iteration Variants
Focused value iteration
41
Value Iteration Variants
Initial Belief
42
Value Iteration Variants
Reachable Beliefs
43
Value Iteration Variants
Relevant Beliefs
44
Value Iteration Variants
45
Value Iteration Variants
Point-based (PBVI) [Pineau et al., 2003]
46
Value Iteration Variants
47
Value Iteration Variants
Heuristic Search (HSVI) [Smith and Simmons, 2004]
48
Outline Motivation Background and related work Technical approach
Preliminary work Proposed research
49
Intelligent Site Survey
50
Intelligent Site Survey
51
Intelligent Site Survey
52
Intelligent Site Survey
!
53
Intelligent Site Survey
54
System Architecture planner executive functional layer environment
55
System Architecture planner executive functional layer science
observer environment
56
System Architecture planner domain planners executive functional layer
science observer environment
57
System Architecture off-board onboard planner domain planners science
interface executive off-board planner functional layer science observer environment
58
Pushing POMDP Technology
Heuristic search Combine heuristic search with efficient value function representations Factored state Do better when the state is mostly observable Data structures that leverage independence Continuous planning Fast replanning
59
Outline Motivation Background and related work Technical approach
Preliminary work Proposed research
60
Heuristic Search Value Iteration
Solves POMDPs Fast anytime algorithm with provable bounds on plan quality On some benchmark problems, >100x speedup V * V
61
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
62
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
63
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
64
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
65
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
66
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
67
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
68
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
69
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
70
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
71
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
72
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
73
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
74
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
75
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
76
Heuristic Search Value Iteration
Focus on reachable beliefs Avoid considering foolish actions and unlikely observations
77
Bounds Update at b b
78
HSVI Performance Tag (870s 5a 30o)
79
Multi-Algorithm Comparison
80
Comparison with PBVI [Pineau et al., 2003]
optimal value HSVI PBVI solution quality wallclock time
81
Comparison with PBVI [Pineau et al., 2003]
optimal value vc HSVI PBVI reported times HSVI PBVI
82
Comparison with PBVI [Pineau et al., 2003]
83
HSVI Related Work
84
Outline Motivation Background and related work Technical approach
Preliminary work Proposed research
85
Research Goals Develop and field test an SA system, includes developing Problem definition and performance criteria Overall architecture Planning module Extend POMDP state-of-art to meet the needs of the SA problem
86
Performance Criteria: SA
Overall performance: did the science team characterize the site accurately? Subgoals with quantitative metrics Coverage, key signatures, anomalies, representative sampling Compare to baseline operational modes
87
Baseline Operational Modes
88
Baseline Operational Modes
directed sampling
89
Baseline Operational Modes
90
Baseline Operational Modes
periodic sampling
91
Performance Criteria: POMDP
Well-established metrics for POMDP planning performance Primarily test in simulation Compare to baseline planning techniques Ignoring informational value (MDP) Greedy or heuristic techniques (e.g. Q- MDP)
92
Life in the Atacama Robotic astrobiology: work under Mars-relevant operational constraints and do meaningful science on Earth.
93
Science Mission The biodiversity and distribution of habitats in the Atacama is not yet measured. Where does life survive and where does it not? Coastal Range Interior Desert
94
Planner Implementations
Atacama 2004 Make it work: deterministic world model, heuristic search, quick development Technology research A series of quickly developed planners that focus on specific POMDP techniques, tested in simulation Atacama 2005 Integrate lessons from earlier versions: optimize for stability, extensibility POMDP technology may not be ready
95
Pushing POMDP Technology
Heuristic search Combine heuristic search with efficient value function representations Factored state Do better when the state is mostly observable Data structures that leverage independence Continuous planning Fast replanning
96
Contributions Lessons learned from developing and field testing one of the first science autonomy systems Distinguished by “keep moving” operational strategy and novel intelligent site survey operational mode Extensions to the POMDP state-of-art that apply both to SA and other domains
97
Schedule POMDP technology research (ongoing)
Initial SA development (Summer 04) LITA Expedition 04 (Fall 04) POMDP SA model development (Spring 05) LITA Expedition 05 (Fall 05) Thesis writing (Spring-Summer 06)
98
Questions?
99
NASA Mars Agenda 2005 Mars Reconnaissance Orbiter
2007 Mars Scout 1: Phoenix Lander 2009 Mobile Lab 2011 Mars Scout 2 2015 Mars Sample Return ... ~2025 Human Landing
100
OASIS Project Broad agreement on goals and approach
For instance, ideas about preferences (key signature, anomaly, etc.) Distinctions of the proposed work Operational strategy: keep moving Kilometer-scale traverses, multiple sites per day Novel intelligent site survey concept
101
LITA Science Autonomy 2004 September 2004 Expedition
Early versions of all modules present Test science observer/planner closed loop in very simple situations Teleoperate rover to simulate SA system Gain experience with architecture and domain modeling Get feedback from science team
102
LITA Science Autonomy 2005 September 2005 Expedition
Test full-up SA operational modes Learn about integration into overall operations strategy Characterize performance
103
Sensing Uncertainty type 4 type 7 novel
104
Sensing Uncertainty type 4 type 7 novel type 4 type 7 novel
105
Sensing Uncertainty type 4 type 7 novel type 4 type 7 novel type 4
106
Sensing Uncertainty type 4 type 7 novel type 4 type 7 novel type 4
107
Science on the Fly
108
Science on the Fly
109
Science on the Fly
110
Science on the Fly
111
Science on the Fly !
112
Science on the Fly
113
Science on the Fly carbonate!
114
Science on the Fly
115
Science on the Fly
116
Science on the Fly
117
Science on the Fly
118
Science on the Fly
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.