Learning Structure, Reusability and Realtime Modeling in Teams of Autonomous Robots

1
Learning Structure, Reusability and Real-time
Modeling in Teams of Autonomous Robots

• Manuela Veloso
• Tucker Balch
• The MultiRobot Lab, CMU

2
Research trusts

• Bridging the gap between low-level and high-level
  control
• Scalable command and control for large numbers of
  autonomous robots
• Real-time adversary modeling
• Reuse of learned subproblems

3
Learning and modeling the world

• Domain Sony legged robot
• Challenges
• Self localization
• Cooperative localization
• Teammate and adversary recognition
• Opponent modeling

4
Probabilistic localization

• Bayesian approaches
• Evidence grids (Elfes 1989)
• Grid-based Bayesian update (Veloso Uther 98)
• Monte Carlo Localization (Fox, et al 1999)
• Advantages of MCL
• Samples are exact points (no grid-based aliasing)
• Localization is point average
• Updates are easier

5
Monte Carlo Localization

• Notation
• P(l) locale probability density
• P(ll,m) movement probability density
• P(ls) sensor reading probability density
• Movement updates
• convolution of the locale probability density
  P(l) with the movement probability density
  P(ll,m)
• Sensor updates
• multiplication of the locale probability density
  P(l) by the sensor reading probability density
  P(ls)

6
Challenges for real-time localization

• Unmodeled terrain may significantly impact
  movement
• Teleportation or large movement errors require
  fast global localization (versus tracking only)
• Low-power CPU
• High variance in movement outcomes
• Low-resolution vision

7
Limitations of MCL

• Requires hundreds of samples for robustness
• Takes too long to handle large errors in movement
  modeling
• Cannot handle unmodeled movements
• Bottom line too slow on Sony platform

8
Solution Sensor Resetting Localization (Lenser
Veloso, ICRA-2000)

• SRL - new probabilistic localization algorithm,
  an extension of MCL
• SRL adds an additional step to the sensor update
  phase of localization
• If the probability of the estimated position is
  low given the sensor readings, SRL replaces some
  samples with samples drawn from the pdf given by
  the sensors (P(ls)).

9
Sensor Resetting Localization
10
Sensor Resetting Localization
11
Sensor Resetting Localization
12
Comparison of SRL with MCL
13
Comparison of SRL with MCL
14
SRL Video

• Shot yesterday

15
Building a large-scale robot team

• Robot platform design
• Low cost
• Rich sensing
• Moderate compute power
• Reliable moderate bandwidth communication
• Software tasks
• TeamBots environment
• Simulation
• Hardware integration

16
Robot platform

• Mechanical platform Cye by Probotics
• Sensing
• Bump sensing
• Odometry
• Color vision _at_ 30Hz
• Communication
• 1mbs wireless ethernet
• Computing
• 300Mhz Pentium/Linux
• Price lt3000

17
Robot platform
18
TeamBots

• Same control system runs in simulation and on
  mobile robots

Control System
API
Simulated Robot
Robot Hardware
19
TeamBots other features

• Java
• Supports multiple mobile robot platforms
• Complex simulation environments described in
  simple human-readable files
• Integrated inter-robot communication package
• Freely available on the web (www.teambots.org)

20
Integrating the new platform with TeamBots

• Simulation
• Combination of diff-drive robot and trailer
• Polygonal
• Mobile robot
• Low-level control API in Java
• TeamBots API
• Integration of vision library
• Video

21
Real-time color image segmentation (Bruce, Balch
Veloso, in submission)
22
Recent relevant publications

• Vision-servoed localization and behavior-based
  planning for a quadruped legged robot, Veloso,
  Winner, Lenser, Bruce Balch, AIPS-2000
• Sensor Resetting Localization for poorly modeled
  mobile robots, Lenser Veloso, ICRA-2000
• Social potentials for scalable multi-robot
  formations, Balch, ICRA-2000
• Real-time color image segmentation using
  commodity hardware, Bruce, Balch Veloso, in
  submission
• On behavior classification in adversarial
  environments, Riley Veloso, in submission

23
Whats next...

• Scale up to 5 (or more) robots
• Domains
• Robotic soccer
• Search and rescue
• Research issues
• Cooperative localization
• Communication hierarchies for strategy and
  shared sensing
• Learning real-time modeling of adversaries
• Real-time multirobot continuous planning

24

• www.cs.cmu.edu/multirobotlab
• www.teambots.org