Model-based Programming of Cooperative Explorers

1
Model-based Programming of Cooperative Explorers

• Prof. Brian C. Williams
• Dept. of Aeronautics and Astronautics
• Artificial Intelligence Labs
• And Space Systems
• MIT

2
Classical Robotics
3
Deep Space Explorers and Aerial Robots
Cassini Saturn Orbital Insertion Complements JPL
Self Repairing
Agile
Human-guided stunt helicopter Complements Eric
Feron, MIT and Draper Labs
4
Cooperative
Exploring Mars Hetrogenous robotics test bed
5
Model-based Autonomy and Programming

• How do we program explorers that reason quickly
  and extensively from commonsense models?
• How do we incorporate model-based autonomy into
  every day, ubiquitous computing devicesthrough
  model-based programming?
• How do we coordinate heterogeneous teams of
  robots -- in space, air and land -- to perform
  complex exploration?
• How do we couple reasoning, adaptivity and
  learning to create robust agents?

6
Model-based Programming ofSelf Repairing
Explorers
7

• Leading Diagnosis
• Legs deployed during descent.
• Noise spike on leg sensors latched by software
  monitors.
• Laser altimeter registers 50ft.
• Begins polling leg monitors to determine touch
  down.
• Latched noise spike read as touchdown.
• Engine shutdown at 50ft.

• Mars 98
• Climate Orbiter
• Mars Polar Lander

8
Monitors
Command Sequence
Mission Goal Scenario
9
Robustness involves controlling and sensing a
complex internal network
10
Managing Interactions for Self-Repair
Programmers and operators must reason through
system-wide interactions to

• monitor
• track goals
• confirm commands
• isolate faults
• diagnose faults

• reconfigure hardware
• coordinate control policies
• repair
• avoiding failures

11
Robustness Requires Novel Solutions

• 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.

courtesy of NASA
12
Example of a Model-based Agent
Goals
Scripts

• Goal-directed
• First time correct
• projective
• reactive
• Commonsense models
• Heavily deductive

Remote Agent
Scripted Executive
Mission Manager
Planner/ Scheduler
Diagnosis Repair
Mission-level actions resources
component models
13
Started January 1996 Launch Fall 1998
courtesy JPL
14
Remote 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

15
Model-based Autonomy and Programming

• How do we program explorers that reason quickly
  and extensively from commonsense models?
• How do we incorporate model-based autonomy into
  every day, ubiquitous computing devicesthrough
  model-based programming?
• How do we coordinate heterogeneous teams of
  robots -- in space, air and land -- to perform
  complex exploration?
• How do we couple reasoning, adaptivity and
  learning to create robust agents?

16
Model-based Programs Interact Directly with State

• Embedded programs interact withplant
  sensors/actuators
• Read sensors
• Set actuators

• Model-based programs interact with plant state
• Read state
• Write state

Programmer must map between state and
sensors/actuators.
Model-based executive maps between sensors,
actuators to states.
17
Orbital Insertion Example
EngineA
EngineB
EngineA
EngineB
Science Camera
Science Camera
18
Model-based Programs Evolve Hidden State
Reactive Control Program
goals
State
Model-based Executive
Command
Observations
19
Model-based Programs Evolve Hidden State

• Reactive Model-based Programming Language
• State assertion
• State query
• Conditional execution
• Preemption
• Iteration
• Concurrent execution

Programmer specifiesabstract state evolutions
Temporal planner
Programs read and set variables that are
typically hidden.
Goal state
Inferred State
Model-based Executive
Command
Observations
20
RMPL Executive Reasons from Plant Model
Temporal planner
goals
State
Model-based Executive
Command
Observations
21
RMPL Executive Reasons from Plant Model
Temporal planner
goals
State
Model-based Executive
mode open ? high pressure high flow
nominal pressure nominal flow . . . mode
closed ? Zero flow
Command
Observations
22
RMPL Executive Reasons from Plant Model
Temporal planner
Achieve Thrust
goals
State
Model-based Executive
Command
Observations
23
Model-based Programming
Automate programmers ability to reason from
commonsense models in light of mission goals.

• Model-based Reactive Programming
• Programmer guides state evolution at strategic
  levels.
• Commonsense Modeling
• Programmer specifies commonsense, compositional
  models of spacecraft behavior.
• Model-based Execution Kernel
• Reason through system interactions on the
  fly,performing deduction within the reactive
  control loop.
• Model-based Compilation
• Reduce knowledge to a vivid form, think through
  local contingencies.

24
OPSAT Exploring the Improbable

• When you have eliminated the impossible,
  whatever remains, however improbable, must be the
  truth. - Sherlock Holmes. The Sign of the Four.

25
Compare Most Likely Hypothesis to Observations
It is most likely that all components are okay,
and the valves are in the mode indicated.
26
Isolate Conflicting Information
The component modes indicated in red conflict
with the model and observations.
27
Leap to the Next Most Likely Hypothesisthat
Resolves the Conflict
The next consistent hypothesis must differ from
the conflict by at least one component mode.
28
New Hypothesis Exposes Additional Conflicts
Next consistent hypothesis must differ from both
conflicts
29
Final Hypothesis Resolves all Conflicts
Helium tank
Fuel tank
Oxidizer tank
Pressure1 nominal Flow1 zero
Pressure2 nominal Flow2 positive
Main Engines
Acceleration zero
Implementation Hypotheses generated by minimal
set covering, guided by A search.
30
Model-based Autonomy and Programming

• How do we program explorers that reason quickly
  and extensively from commonsense models?
• How do we incorporate model-based autonomy into
  every day, ubiquitous computing devicesthrough
  model-based programming?
• How do we coordinate heterogeneous teams of
  robots -- in space, air and land -- to perform
  complex exploration?
• How do we couple reasoning, adaptivity and
  learning to create robust agents?

31
Model-based Programming of Cooperative Systems

• How do we model behaviors of cooperative robotic
  networks?
• How do we command cooperative networks?
• How do we monitor cooperative networks?

32
Managing Interactions for Cooperation
Programmers and operators must reason through
system-wide interactions to

• select among redundant procedures
• Evaluate outcomes
• Plan contingencies

• select deadlines
• select timing constraints
• allocate resources

33
Model-based Cooperative Programming

• Model-based Programs
• Specify team behaviors as concurrent programs.
• Specify alternatives using decision theoretic
  choice.
• Specify timing constraints between activities.
• Model-based Execution
• Looks before it leaps
• Pre-plans threads of execution that are
  optimal and temporally consistent.
• Responds at reactive timescales
• Perform planning as graph search
• Navigates globally

34
Mission Scenario
ONE
TWO
HOME
Enroute
RENDEZVOUS
RESCUE AREA
Station ABC
Diverge
Station XYZ
RESCUE LOCATION
MEETING POINT
35
Enroute Activity
Enroute
Corridor 2
Rendezvous
Rescue Area
Corridor 1
Corridor 3
36
RMPL for Group-Enroute
Non-deterministic choice

• Group-Enroute()l,u
• choose
• do
• Group-Traverse-Path(PATH1_1,PATH1_2,PATH1
  _3,RE_POS)l90,u90
• maintaining PATH1_OK,
• do
• Group-Traverse-Path(PATH2_1,PATH2_2,PATH2
  _3,RE_POS)l90,u90
• maintaining PATH2_OK
• Group-Transmit(OPS,ARRIVED)0,2,
• do
• Group-Wait(HOLD1,HOLD2)0,u10
• watching PROCEED

37
RMPL Interpreter

• Dynamically selects among alternate executions,
  satisfies open conditions and checks
  schedulability,
• Selects execution times, monitors outcomes and
  plans contingencies.

• Kirk
• Idea
• Titan

38
Enroute Activity

• Least cost threads of execution generated by
  extended auction algorithm

price 425
price 0
1
2
450,540
0
0
Extend Path
4
5
9
10
425
30
0
0
0
0
price 425
price 0
price 30
price 0
8
13
3
price 0
price 0
0
price 425
0
0
0
6
7
11
12
440
1
price 440
price 0
price 1
price 0
9 ? 10
?
?
13 ? 2
Path P 1 ? 3 ? 4 ? 5 ? 8
?
?
11 ? 12
Start Node 1 End Node 2
39
Temporal planning is combined with randomized
path planning to find a collision free corridor
Path 1
xinit
xgoal
Xobs
4
5
40
Robotic Vehicles

• ATRV Rovers
• Monster Trucks
• Blimps
• Spheres
• Simulated Air/Space Vehicles

41
Scenario
Cooperative Target Site Exploration Heterogeneous
rover team and blimps explore science sites
determined by remote sensing

• Tasks
• small scout rovers (ATRV Jr) explore terrain as
  described in earlier scenarios
• blimps provide additional fine grain air
  surveillance
• scout rovers identify features for further
  investigation by sample rover (ATRV)
• scout rovers provide refined terrain mapping for
  path planning of the larger sample rover
• Scenario Research Objective
• Extend coordination to heterogeneous team

exploration region identified feature goal
position
42
Summary Model-based Autonomy

• To survive decades autonomous systems must
  orchestrate complex networks of devices.
• In model-based autonomy, programming involves
  specifying strategies as abstract state
  evolutions, and specifying commonsense models.
• Runtime kernels are agile, deducing and planning
  from these models within the reactive loop.
• Model-based programming languages and kernels are
  broadening out to cooperative and hybrid systems.