Adaptive Coordinated Control of Intelligent MultiAgent Teams

1
Adaptive Coordinated Control of Intelligent
Multi-Agent Teams

• Vijay Kumar
• Thrust II Lead

2
Thrust II

• Acquisition and Integration of Rich Multi-sensor
  Information into Virtual Environments for
  Incorporating Human Intervention in Mission
  Planning and Execution
• Acquisition of multi-sensor information using
  multiple mobile sensory agents
• Adaptive hierarchical networks for acquiring and
  providing information
• Extraction of 3D models from distributed sensors
  networks and
• Environments for human intervention and decision
  making.

3
Keys to Thrust II

• Network of sensors, actuators, and controllers
• Adaptive
• To who can see who, who can talk to who
• Integrative
• From isolated 1-D or 2-D measurements to
  continuous space-time model
• Transparent
• To specifics of platform, who is doing what
• Proactive
• Mobility in support of communication and
  perception
• Scalable
• Distributed sensing and computation

4
Thrust II Presentations and Demonstrations

• Acquisition of multi-sensor information using
  multiple mobile sensory agents
• Ben Grocholsky, Ethan Stump and Nathan Michael
• Ben Grocholsky
• Adaptive hierarchical networks for acquiring and
  providing information
• Ani Hsieh and Anthony Cowley
• Extraction of 3D models from distributed sensors
  networks
• Kostas Danillidis
• C. J. Taylor
• Environments for human intervention and decision
  making
• C. J. Taylor and Anthony Cowley

• Sandy Patterson and Ameesh Makadia
• Arvind Bhusmurnath and Babak Shir

5
Coordination between MURI participants

• Network of vehicles for situational awareness
  Taylor, Kumar
• Omnidirectional cameras, vision-based control
  Bajcsy, Danillidis, Geyer, Isler, Sastry
  Air-ground coordination for search, detection and
  localization Kumar, Pappas, Taylor
• Control/planning of undulatory locomotion systems
  Choset, Kumar

6
Technology Transfer

• Acquisition of multi-sensor information using
  multiple mobile sensory agents
• Lockheed Martin
• Adaptive hierarchical networks for acquiring and
  providing information
• Boeing, Lockheed Martin
• Extraction of 3D models from distributed sensors
  networks and
• Adaptive Threat Detection
• Environments for human intervention and decision
  making
• Microsoft
• Interest from the Navy, NIST

ARL DARPA ONR Navy DEMO
7
Penn Acclimate Alumni

• Chris Geyer (CMU)
• Volkan Isler (RPI)
• Bert Tanner (UNM)
• Paolo Tabuada (Notre Dame)
• Calin Belta (BU)
• John Spletzer (Lehigh)
• Rahul Rao (Intel)

8
This Talk

• Demonstration at Fort Benning (DVD)
• Active Localization of Robots and/or Targets (Ben
  Grocholsky and Ethan Stump)
• Deployment of Networked Vehicles (Ani Hsieh and
  Anthony Cowley)
• Verification and Falsification of
  Plans/Strategies for Multi-Vehicle Systems (Peng
  Cheng and Jongwoo Kim)

9
Active Localization

• Move to localize

Where and how the robot/sensor moves affects the
measurement and the estimate
10
The Next Best View Problem

• Move to maximize expected information gain

11
Cooperative search, identification, and
localization
12
Information Model Shared
13
Confidence Ellipsoids


14
UGV Trajectory
15
UAV search pattern
UGV identification and localization of potential
targets
Grocholsky et al, ISER, 2004 Grocholsky et al,
ICRA, 2005
16
Multiple Vehicles, Multiple Targets
17
Set-Valued Representations

• Motivation
• Avoid models of measurement noise
• Avoid linearization based approaches
• Approach
• Use set-valued models of measurement errors
• Measurements impose linear constraints in a
  higher dimensional space
• Ellipsoidal calculus gives us powerful tools for
• Propagation, update, fusion, intersection,

18
Example

• Localization of one vehicle with range sensor

19
The Higher Dimensional Space

• Observation model is linear if x, y, x2y2 is
  used as a basis

x, y
xL, yL
20
Two measurements of the same landmark
21
Localizing Four Robots with Range Sensors
Basis
Set valued representation of uncertainty
Steps

• Ellipsoidal representations for uncertainties in
  higher-dimensional space
• Intersect ellipsoids to fuse information
• Project ellipsoids to obtain uncertainty in
  lower-dimensional (real) space

22
Localizing Four Robots with Range Sensors
23
Optimal Control for Localization
Robot
Target
24
Indoor Testbed
Funded by ARO W911NF-04-1-0148 (DURIP)
25
Deployment of Mobile Robot Network with
End-to-End Performance Guarantees

• Ani Hsieh and Anthony Cowley

26
Radio Mapping in UrbanEnvironments

• Radio signal strength (s/n) varies across site
• Mapping instead of modeling

M. A. Hsieh, V. Kumar, and C. J. Taylor,
Constructing Radio Signal Strength Maps with
Multiple Robots, IEEE International Conference
on Robotics and Automation, New Orleans, LA,
April 2004.
27
Operator Interface
28
Team Size Affects Available Bandwidth
1, 5
1, 2, 5
No. Transactions per unit time
1, 2, 3, 5
1, 2, 3, 4, 5
29
Signal Strength vs. Bandwidth
Normalized signal strength
30
Link Quality Constrained Navigation

• if LinkQuality lt Minimum then
• Recover
• end if
• if Minimum lt LinkQuality Acceptable then
• if Recovering then
• Stop and wait
• end if
• if waitTime gt MaxWaitTime then
• Retry going to goal
• end if
• end if
• if LinkQuality gt Acceptable, then
• Go to goal
• end if

31
Two Types of Experiments

• Changes in Team Size
• Path graph

• Perimeter Surveillance
• Star graph

32
Changes in Team Size
33
Perimeter Surveillance(Monitoring Signal
Strength)
34
Perimeter Surveillance(Monitoring Target
Transaction Rates)
35
Network Centric Network Sensitive
36
Network Centric Network Sensitive
M. A. Hsieh, A. Cowley, V. Kumar, and C. J.
Taylor, Deploying Networked Robots, Technical
Report, GRASP Laboratory, August 2005 (submitted
to ICRA 2006).
37
Verification and Falsification of Plans for
Unmanned Vehicles

• Peng Cheng, Jongwoo Kim, Vijay Kumar

38
Verification and Falsification of Plans for
Unmanned Vehicles
Surveillance Point

• Blue Team trajectory xb(t)
• Red Team trajectory xr(t)
• Unsafe Region for Blue Team
• g(xb(t), xr(t)) 0
• Falsification find xr(t) such that
• g(xb(t), xr(t)) 0

UGV in Red Team
UGV Motion Range
UGV Detection Range
UAV in Blue Team
Unsafe Region

• Verification ensure that there is no xr(t) such
  that
• g(xb(t), xr(t)) 0

39
Sampling-Based Algorithms

• Exact and finite representations of the search
  space of general verification problems are not
  available.
• The seach space is approximated using sampled
  points.

• Iteratively build search graph G(N,E) and
  construct xr(t) with sampled controls and states

40
Algorithm for Finding Falsifying Control Inputs

• Falsification given e f gt 0, find xr(t) such
  that
• g(xb(t), xr(t)) - e f

• Generate a finite number of sampled controls
  according to given e f .
• Use sampled controls to construct search graph
  and falsifying controls.
• Heuristic in generating search graph.

41
Verification Algorithm

• Resolution completeness of sampling-based
  falsification algorithm

Unsafe Region
If a desired falsifying control with the
specified clearance exists, a resolution complete
sampling-based algorithm will find its
approximation in finite time.
e v
Trajectories constructed by sampling-based
falsification

• Blue Team is verified if a resolution complete
  sampling-based falsification algorithm fails to
  find xr(t) such that
• g(xb(t), xr(t)) e v

42
Controls of Unconventional Locomotion Systems

• Sachin Chitta, Peng Cheng, Vijay Kumar

43
Control and Planning for RoboTrikke,
RollerBlader, and RollerRacer

• Visual feedback based control for RoboTrikke
• Chitta, Cheng, Frazzoli, Kumar, ICRA 2005
• Motion planning using composition of gaits for
  RollerBlader Chitta, Kumar, DETC 2004
• Motion planning for a novel sticking/slipping
  switching model for RollerRacer Cheng, Frazzoli,
  Kumar, 2006, submitted

RoboTrikke
RollerBlader
RollerRacer