Integrated, PlanBased Control of Autonomous Robots in Human Environments

1
Integrated, Plan-Based Control of Autonomous
Robots in Human Environments

• Presented by Aaron Drew
• April 17, 2003
• Michael Beetz, Munich University of Technology
• Tom Arbunckle, Thorsten Belker, Armin B. Cremers,
  Dirk Schulz, University of Bonn
• Maren Bennewitz, Wolfram Burgard, Dirk H hnel,
  University of Freiberg
• Dieter Fox, University of Washington
• Henrik Grosskreutz, Aachen University of
  Technology

2
The Challenge

• The next step in autonomous robots.
• Create a robot which has several functions
• Operates in an unstructured, dynamic environment.
• Accomplishes prolonged, complex, and dynamicly
  changing tasks.
• Performs safely and reliably.

3
The Requirements

• The robot must possess several abilities
• Rich perceptual capabilities to recognize objects
  and places.
• Method to communicate with people.
• System to manage concurrent tasks.
• Self improving control routines based on
  experience.

4
The Star of Our Show

• Rhino, a RWI B21 mobile robot.
• Has a unique control system.
• Enables successful operation in increasingly
  complex environments.
• Has multiple vision sensors.
• Includes human interaction functions.

5
Dynamic-system Perspective

• Must be flexible and responsive to changes.
• Dynamic systems used as primary abstract model
  for programming integrated, plan-based
  controller.
• Two processes
• Controlled process
• Controlling process

6
Controlled Process

• Events in environment, robot's physical
  movements, and sensing operations.
• Broken into two processes
• Environment
• Changes world state.
• Sensing
• Maps world state into sensor data.

7
Controlling Process

• Robot's controlling system.
• Broken into two processes
• State estimation
• Computes beliefs about the controlled system's
  states.
• Action generation
• Specifies control signals supplied to controlled
  process as a response to estimated system state.

8
Visualization of Processes
9
RPL

• Uses Reactive Plan Language.
• Provides traditional control abstractions.
• Includes high-level constructs (interrupts,
  monitors).
• Allows specification of interactive behavior,
  concurrent control processes, failure recovery
  methods, and temporal coordination of processes.
• Makes plans reactive and robust by incorporating
  sensing and monitoring, as well as reactions.

10
Plan-based High-level Control

• Robot must flexibly interleave tasks, exploit
  opportunities, quickly plan actions, and revise
  intended activities.
• Rhino plans represent future activity and have
  two roles
• Executable prescriptions the robot interprets to
  accomplish jobs.
• Syntactic objects the robots revises to meet
  specific success criteria.

11
Planning Processes

• Project what might happen when a robot controller
  executes a plan.
• Return the result as an execution scenario.
• Infer what might be wrong with a robot controller
  given an execution scenario.
• Perform complex revisions on robot controllers.

12
Probabilistic State Estimation

• Sensors are inaccurate.
• Combine data to get object's most likely state.
• When asked, state estimators return all available
  information in probability density.
• Allows variance and entropy to be measured.

13
Plan Transformation

• Robot has a plan library of canned plans.
• Not always optimal.
• If unexpected opportunity presents itself, canned
  plans will miss them.
• Self-adapting plans are used.
• When a belief changes, a runtime plan adaption
  process occurs.
• Decides if revisions are necessary and if so,
  performs them.

14
How it works

• Move red letter from A-111 to A-120
• Move green book from A-110 to A-113
• Environment stays static.

15
How it works

• Move red letter from A-111 to A-120
• Move green book from A-110 to A-113
• Door to A-120 closes after getting the red letter.

16
Control System Architecture

• Three parts
• Low-level control system
• Responsible for sensing and navigating.
• High-level controller, or structured reactive
  controller
• Must effectively combine tasks into coherent
  task-directed behavior.
• High-level interface
• Let's low-level and high-level talk.

17
System Architecture Visualization
18
Low-Level Control Systme

• 20 distributed modules monitor or control a
  dedicated aspect of the robot.
• Modules communicate using asynchronous
  message-passing library.
• They use iterative algorithms with simple update
  rules.
• Get immediate answer and iteratively improve it.

19
Perceptual Subsystem

• Original Rhino had localization and mapping
  probabilistic state estimators.
• Added two more to allow for a more dynamic
  environent
• One monitored doors, chairs, and waste baskets.
• Other tracked people and other robots.

20
Perceptual Subsystem (2)

• Also an image-processing subsystem.
• Uses RECIPE (reconfigurable, extensible, capture
  and image processing environment).
• Loads modules for image processing at runtime.

21
Perceptual Subsystem (3)

• Another component supports some natural language.
• Rhino able to be instructed by emails.
• Supports only a constrained subset of English.
• Parsed with a simple definite-clause grammar.

22
Perceptual Subsystem (4)

• From peters_at_cs.uni-bonn.de
• Date Fri, 24 Oct 1997 120357
• To rhinotcx_at_cs.uni.bonn.de
• Subject Command
• Could you please bring the yellow book on the
  desk in room a-120 to the library before 1230

23
Perceptual Subsystem (5)

• (Request
• SENDER (THE PERSON (FIRST-NAME Hanno)
• (LAST-NAME Peters))
• RECEIVER (THE ROBOT (NAME RHINO))
• TIME (THE TIME-INSTANT 10241203)
• REPLY-WITH (YOUR COMMAND FROM 1203)
• CONTENT (ACHIEVE
• (LOC (THE BOOK
• (COLOR YELLOW)
• (ON (THE DESK
• (IN (THE ROOM A-120)))))
• (THE LOC (IN (THE ROOM LIBRARY))))
• DEADLINE (A TIME INSTANT (BEFORE (DATE 1230)))

24
Navigational Subsystem

• Transforms a pair of locations into a Markov
  Decision Process.
• Solves using a value iterator.
• End result is a mapping from every possible
  location to a optimal heading for the
  destination.
• Reactive collision avoidance module gets mapping.
• Mapping, dynamics, and sensor data used to reach
  to reach destination.

25
High-Level Interface

• Provides a mechanism for High-Level Controller to
  control low-level control processes.
• High-Level Controller can start, terminate, and
  wait for processes.
• Interface gives High-Level Controller feedback,
  such as task completion signals.

26
Structured Reactive Controllers

• The high-level controller.
• Given a set of jobs, SRC concurrently executes
  default routines.
• During the execution of routine activities, SRC
  determines if plans will interfere with each
  other and monitors if robot experiences
  non-standard situations.
• If necessary, determines how routine may work and
  revises plan to make it more robust.

27
SRC Prediction

• Rhino's action model reflects several facts
• Physical robot actions cause continuous change.
• Controllers are reactive systems.
• The robot is executing multiple physical and
  sensing actions.
• The robot is uncertain about the effects of its
  actions and the state of the environment.
• Huge branching factor for future possible states.

28
SRC Prediction (2)

• Rhino employs probabilistic sampling-based
  temporal projection methods.
• Infer information quickly with a bounded risk.
• Given a set of possible plan failure modes and a
  risk the robot is willing to take that its
  prediction is wrong, infer whether the plan is
  likely to produce a failure mode with a
  probability greater than a given threshhold.

29
Transformational Planning of Concurrent Reactive
Plans

• Simple version take several plans, glue them
  together, improve the plan so it works.
• Complex version
• Search in plan space. Nodes are proposed plans.
• Initial node is default plan from plan library.
• Adaptor projects plan to find execution
  scenarios.
• Adaptor criticizes execution scenarios to
  estimate how good a plan is and predict possible
  plan failures.
• Adaptor revises plan to produce a new version,
  and repeats.

30
Learning Symbolic Robot Plans

• XFRMLEARN.
• Starts with default plans made through navigation
  system and MDP.
• Activates collision avoidance module and
  executes.
• XFRMLEARN watches the behavior, and finds
  stretches to improve (thin hallways, for
  example).
• Forms revision methods and experiments with them.
  Successful methods are kept and used.

31
Sample of Navigation Learning
32
Future Work

• Improving robustness for hardware and software
  failures, along with unknown situations.
• Extend probabilistic models and context-sensitive
  calculation abilities.