Title: Robotic Space Explorers
1Robotic Space Explorers
- Brian C. Williams
- Space Systems Lab
- Artificial Intelligence Lab, MIT
2Marskokhod at NASA Ames research center
3Smart Buildings at CMU Xerox PARC
4Ecological Life SupportFor Mars Exploration
5Portable Satellite Assistant
Courtesy of Yuri Gawdiak, NASA Ames
6MIT SSL
courtesy Dave Miller, SSL MIT
7Robotic Webs
8Unmanned Air Vehicles
9Intelligence Embedded at all Levels
- Behavior-based robotics Subsumption
- Reinforcement learning and MDPS
- Classical planning and execution
- Model-based diagnosis and execution
- Mission-level planning
- Robotic path planning
- Probabilistic monitoring and
- Decision-theoretic planning
- Multi-agent coordination
Increased Reasoning
10To Boldly Go Where No AI System Has Gone Before
11Started January 1996 Launch Fall 1998
courtesy JPL
12Remote Agent Team Members
- Douglas Bernard JPL
- Steve Chien JPL
- Greg Dorais Ames
- Julia Dunphy JPL
- Dan Dvorak JPL
- Chuck Fry Ames
- Ed Gamble JPL
- Erann Gat JPL
- Othar Hansson Thinkbank
- Jordan Hayes Thinkbank
- Bob Kanefsky Ames
- Ron Keesing Ames
- James Kurien Ames
- Bill Millar Ames
- Sunil Mohan Formida
- Paul Morris Ames
- Nicola Muscettola Ames
- Pandurang Nayak Ames
- Barney Pell Ames
- Chris Plaunt Apple
- Gregg Rabideau JPL
- Kanna Rajan Ames
- Nicolas Rouquette JPL
- Scott Sawyer LMMS
- Rob Sherwood JPL
- Reid Simmons CMU
- Ben Smith JPL
- Will Taylor Ames
- Hans Thomas Ames
- Michael Wagner 4th Planet
- Greg Whelan CMU
- Brian C. Williams Ames
- David Yan Stanford
13I am a HAL 9000 computer production number three.
I became operational at the H.A.L. plant in
Urbana, Illinois on January 12, 1997.
14courtesy NASA
International Space Station 1998-2002
15Our vision in NASA is to open the Space
Frontier. When people think of space, they
think of rocket plumes and the space shuttle.
But the future of space is in information
technology. We must establish a virtual
presence, in space, on planets, in aircraft and
spacecraft. - Daniel S. Goldin, NASA
Administrator Sacramento, California, May 29, 1996
- Motive Astrobiology Origins Programs
- Means New Millennium Program
- Smarts Autonomous Reasoning
16Motive Primitive life on Early Mars?
1997 Mars Pathfinder and Sojourner
courtesy JPL
17New Means
Mars Pathfinder, 1997
courtesy JPL
18New Means Mars Airplane
courtesy NASA Ames
courtesy NASA Lewis
19Cryobot Hydrobot
Motive life under Europa?
courtesy JPL
20Formation Flying Optical Interferometer (ST3)
Motive Earth-like Planets Around Other Stars?
courtesy JPL
21Four launches in 7 months
Mars Climate Orbiter 12/11/98
Mars Polar Lander 1/3/99
QuickSCAT 6/19/98
Stardust 2/7/99
courtesy of JPL
22Traditional spacecraft commanding
23How Will They Survive?
- Vanished
- Mars Observer
- Mars Polar Lander
- Miscommanded
- Clementine
- Mars Climate Orbiter
courtesy of JPL
24STS-93 Hydrogen Leak
- Symptoms
- Engine temp sensor high
- LOX level low
- GNC detects low thrust
- H2 level low (???)
- Problem Liquid hydrogen leak
- Effect
- LH2 used to cool engine
- Engine runs hot
- Consumes more LOX
25Intelligent Embedded Systems Cassini
- 7 year cruise
- 150 - 300 ground operators
- 1 billion
- 7 years to build
Faster, Better, Cheaper
- 150 million
- 2 year build
- 0 ground ops
Cassini Maps Titan
courtesy JPL
26Ames-JPL NewMaap New Millennium Advanced
Autonomy Prototype
- no Earth Comm
- 1 hr insertion window
- engines idle for several years
- moves through ring plane
July - November, 1995
courtesy JPL
27(No Transcript)
28Reconfiguring for a Failed Engine
Fuel tank
Oxidizer tank
Open four valves
Valve fails stuck closed
Fire backup engine
29AI in the pre-90s Reservations about Embedded
Systems being Intelligent
- For reactive systems proving theorems is out
of the question Agre Chapman 87 - Diagnostic reasoning from a tractable model is
largely well understood. However we dont know
how to model complex behavior... Davis
Hamscher 88 - Commonsense equations are far too general for
practical use. Sacks Doyle 91
30Houston, We have a problem ...
- Quintuple fault occurs (three shorts, tank-line
and pressure jacket burst, panel flies off). - Mattingly works in ground simulator to identify
new sequence handling severe power limitations. - Mattingly identifies novel reconfiguration,
exploiting LEM batteries for power. - Swaggert Lovell work on Apollo 13 emergency rig
lithium hydroxide unit.
courtesy of NASA
31Challenge Thinking Through Interactions
Programmers must reason through system-wide
interactions to generate codes for
- command confirmation
- goal tracking
- detecting anomalies
- isolating faults
- diagnosing causes
- hardware reconfig
- fault recovery
- safing
- fault avoidance
- control coordination
Equally problematic at mission operations level
32Model-based Autonomy
- Programmers generate breadth of functions from
commonsense models in light of mission goals. - Model-based Programming
- Program by specifying commonsense, compositional
declarative models. - Model-directed Planning Execution
- Provide services that reason through each type of
system interaction from models. - on the fly reasoning requires significant search
deduction within the reactive control loop.
33Towards a Unified Model
Mission Operations Model
Hardware Commanding Failure Model
- Transition Systems Constraints Probabilities
34Fast Search Deep Blue beats Kasparov by brute
force.
35Many problems arent so hard
1
0
0
5
0
1
4
0
2
4
P
h
a
s
e
T
r
a
n
s
i
t
i
o
n
f
o
r
3
-
S
A
T
,
N
1
2
t
o
1
0
0
D
a
t
a
R
e
s
c
a
l
e
d
U
s
i
n
g
4
.
1
7
,
1
.
5
c
a
n
(
K
i
r
k
p
a
t
r
i
c
k
a
n
d
S
e
l
m
a
n
,
S
c
i
e
n
c
e
,
M
a
y
1
9
9
4
)
36General Deduction Can Achieve Reactive Time
Scales
Many problems arent so hard
RISC-like, deductive kernel
4-sat cost
Agenda
TMS
25 var.
FOUND UNSOLVABLE
SOLUTION FOUND
generate successor
conflict database
Average constraints per variable
37Services For Thinking Through Interactions
- Quintuple fault occurs (three shorts, tank-line
and pressure jacket burst, panel flies off). - Mattingly works in ground simulator to identify
new sequence handling severe power limitations. - Mattingly identifies novel reconfiguration,
exploiting LEM batteries for power. - Swaggert Lovell work on Apollo 13 emergency rig
lithium hydroxide unit.
- Multiple fault diagnosis of unexperienced
failures. - Mission planning and scheduling
- Hardware reconfiguration
- Scripted execution
38Remote Agent Architecture
Ground System
RAX_START
RAX_START
Real-Time Execution
RAX Manager
Flight H/W
Fault Monitors
Planning Experts (incl. Navigation)
39Ames-JPL NewMaap New Millennium Advanced
Autonomy Prototype
July - November, 1995
courtesy JPL
40Mission Manager Sets Goals
Thrust Goals
Power
Attitude
Engine
Off
Off
41Plan!
Thrust Goals
Power
Attitude
Thrust (b, 200)
Engine
Off
Warm Up
42Planner Models
- Objects
- state-variables
- tokens
- Constraints
- compatibilities
- functional dependencies
43Compatibility
Thrust Goals
Power
contains
Attitude
Thrust (b, 200)
Engine
44Compatibility
Thrust Goals
Power
equals
contained_by
Point(b)
Attitude
contained_by
meets
met_by
Thrust (b, 200)
Engine
Warm Up
45Planning/Scheduling Cycle
PLAN
NO
Uninstantiated compatibility
. . .
Instantiate compatibility
Heuristics
. . .
Backtrack
Schedule token
NO
YES
46Types of Plan Flaws
- Un-instantiated compatibilities
- subgoaling
- Un-inserted tokens
- Under-constrained parameter
- Gaps in scheduling horizon
47Plan Generates A Simple Temporal Constraint
Network
????????????
????????
????????
?????????
??????
??????
?????????
48DS1 Planner/Scheduler summary
- Model size (Remote Agent Experiment)
- state variables 18
- token literal types 42
- compatibility specs 46
- Plan size
- tokens 154
- temporal relations 180
- variables 288 (81 time points)
- constraints 232 (114 distance bounds)
- Performance
- search nodes 649
- search efficiency 64
49Executing Temporal Plans
- Propagate time
- Select enabled events
- Terminate preceding tokens
- Run next tokens
50Time Propagation Can Be Costly
EXECUTIVE
CONTROLLED SYSTEM
51Compile to Efficient Network
EXECUTIVE
CONTROLLED SYSTEM
52Model-based Execution of Tokens
Programmers and operators must reason through
system-wide interactions to generate codes for
- monitoring
- tracking goals
- confirming commands
- detecting anomalies
- diagnosing faults
- reconfiguring hardware
- coordinating control policies
- recovering from faults
- avoiding failures
Identifying Modes
Reconfiguring Modes
53Model-based Execution asStochastic Optimal
Control
Model
Goals
mode reconfiguration
s(t)
Controller
mode identification
?(t)
o(t)
Plant
s (t)
f
g
Livingstone
54Models
- modes engage physical processes
- probabilistic automata for dynamics
vlvstuck open gt Outflow Mz(inflow)
vlvopen gt Outflow Mz(inflow)
Vlv closed gt Outflow 0
vlvstuck closedgt Outflow 0
55Model-based programming language
- (defcomponent valve (?name)
- attributes ((sign-values (flow (input ?name)))
- (sign-values (flow (output ?name)))
- ...)
- ...
- (closed
- model (and ( (flow (input ?name)) zero)
- ( (flow (output ?name)) zero))
- transitions ((open-valve when
(open (cmd-in ?name)) next open) - (otherwise persist)))
- ...)
56Mode Identification and Diagnosis
Observe no thrust
Find most likely reachable states consistent
with observations.
57Mode Reconfiguration
Goal Achieve Thrust
58Models Compile to Propositional Logic
- Specifying Variables
- mode ranges over open, closed, stuck-open,
stuck-closed - cmd ranges over open, close, no-cmd
- fin, and fout range over positive, negative,
zero - pin, and pout ranging over high, low,
nominal - Specifying Mode Behaviors
- mode open ? (pin pout) ? (fin fout)
- mode closed ? (fin zero) ? (fout zero)
- mode stuck-open ? (pin pout) ? (fin fout)
- mode stuck-closed ? (fin zero) ? (fout
zero)
59- Specifying nominal transitions
- mode closed ? cmd open ? next (mode open)
- mode closed ? cmd ? open ? next (mode closed)
- mode open ? cmd close ? next (mode closed)
- mode open ? cmd ? close ? next (mode open)
- mode stuck-open ? next (mode stuck-open)
- mode stuck-closed ? next (mode stuck-closed)
- Specifying failure transitions
- ?1 mode closed ? next (mode stuck-closed)
- ?2 mode closed ? next (mode stuck-open)
- ?3 mode open ? next (mode stuck-open)
- ?4 mode open ? next (mode stuck-closed)
60Mode identification and reconfiguration performed
by OPSAT
Combinatorial optimization w propositional
constraints A Conflict-directed Search
DPLL TMS
Check Constraints
Best-first Agenda
Optimal feasible modes
DPLL SAT With ITMS
Checked modes
generate best implicants
Conflicts (infeasible modes)
conflict database
61MI and MR performance
Number of components 80 Number of clauses 11101
LTMS
no TMS
62Started January 1996 Launch Fall 1998
courtesy JPL
63Remote Agent Experiment
See rax.arc.nasa.gov
- May 17-18th experiment
- Generate plan for course correction and thrust
- Diagnose camera as stuck on
- Power constraints violated, abort current plan
and replan - Perform optical navigation
- Perform ion propulsion thrust
- May 21th experiment.
- Diagnose faulty device and
- Repair by issuing reset.
- Diagnose switch sensor failure.
- Determine harmless, and continue plan.
- Diagnose thruster stuck closed and
- Repair by switching to alternate method of
thrusting. - Back to back planning
64Intelligent embedded systems research has
addressed each of AIs concerns
- For reactive systems proving theorems is out
of the question Agre Chapman 87 - Diagnostic reasoning from a tractable model is
largely well understood. However we dont know
how to model complex behavior... Davis
Hamscher 88 - Commonsense equations are far too general for
practical use. Sacks Doyle 91
65 With autonomy we declare that no sphere is
off limits. We will send our spacecraft to
search beyond the horizon, accepting that we
cannot directly control them, and relying on them
to tell the tale. Bob Rasmussen Architect JPL
Mission Data System