Title: DARPA Organic Air Vehicle Project
1DARPA Organic Air Vehicle Project
OAV System Concept
Hand Held Controller
Ground Station
Air Vehicle
Integrated Guidance, Navigation, Control Imaging
2UAV 3D Navigation and Mapping
- Presented AINS C3UV Kickoff MeetingRaja
Sengupta (lead), Zu Kim, Jitendra MalikJoao
Sousa, Sivakumar Rathinam, Marco Zennaro
3Scaling up to Swarms
- Algorithmic Complexity
- Caused by Constrained Resource Allocation
- Appears in the MILP formulation
- How reference..
- Control complexity
- Caused by rate of change of information,
resources, constraints, .
4Scaling up to SwarmsAlgorithmic Complexity
- There are two hopes
- The phase transition literature shows one may
encounter it rarely - Deal with complexity by genetic programming or
simple heuristics - Create complex emergent behaviors by simple rules
5Scaling up to SwarmsAlgorithmic Complexity
- There are two hopes
- The phase transition literature shows one may
encounter it rarely - Deal with complexity by genetic programming or
simple heuristics - Create complex emergent behaviors by simple rules
- Deal with it by statistical or fluid mechanical
abstraction to reduce large numbers - Solve smaller resource allocation problems on the
abstraction
6Scaling up to SwarmsHeuristic Research at UCB
- Collaborative Multi-Vehicle task completion
destination
task
obstacle
obstacle
7Scaling up to SwarmsHeuristic Research at UCB
- Collaborative Multi-Vehicle task completion
8Scaling up to SwarmsHeuristic Research at UCB
- Collaborative Multi-Vehicle task completion
9Scaling up to SwarmsGenetic Programming at UCB
- Swarm navigation through obstacles
10(No Transcript)
11Scalability
- Incremental growth of service network
- add a program or add a vehicle
- SNP integrates it with the rest when required
- Asynchronous
- Incremental growth of management overhead
- search complexity of the Broker increases
gradually with the number of registered services - NP-complete in the worst case sense
Phase transitions have been found for
satistfiability problems
12Scaling up to SwarmsControl Complexity
- Caused by
- Changing structure
- Resources
- Connections
13Scaling up to SwarmsControl Complexity
- Caused by
- Changing information
- Tasks
- Obstacles
- Teammates,
14Scaling up to SwarmsControl Complexity
- Our Innovations
- Service Networks
- Distributed Publish Subscribe
- in multi-vehicle control
- Aim is to control a system of changing dimension
- Keep operational properties invariant even as
system dimension changes - Examples of operational properties
- Agents may leave the system but information does
not (remain invariant) - Each user task will be serviced
15Scalable Information ManagementSearching for
Unknown Threats
16Scalable Information ManagementPublish and
Subscribe
Team Level
Operation decomposer
Information state
Resource allocation/scheduling
Target Distribution Map
Operation monitor
Risk Map
Dispatcher
Formation Navigation
17Scalable Information ManagementTarget Map and
Risk Map
- Target distribution map
- P(A, N, t) probability of N targets of type t in
area A - Target distribution update
- Fuses measurements from different kinds of
sensors (SAR and EO) - Bayesian update
- Risk map computation
- Integral of threat model with respect to the
measure P(A, N, t) - Generates the value function for navigation
18Scalable Information ManagementSearching for
Unknown Threats
19Scalable Information ManagementDistributing the
Publisher Service
- Geographic Data Management Network
20Scalable Information ManagementDistributing the
Publisher Service
User
21Scalable Information ManagementDistributing the
Publisher Service
Movie of our Implementation 4 servers on 4
laptops over wireless
22Controlling Systems with Changing Dimension
- Service Networking Based on Jini Middleware
23What is Jini?
- Service and User arrive at different times
- Easy repair
- Proxy implementation invisible to user
- Easy system integration
Registration
Service
Lookup manager
Multicast Network
RMI
Proxy
ServiceItem
Client
24Systems with Changing DimensionService Network
for Multi-Vehicle Search
Lookup Manager
Mission Control
Search Software
Host
Vehicle
Txn Manager
Search Server
Publisher
25The SN Synthesis Theorem
-
- Theorem Let the set of concrete values be
finite. Then there exists a synthesis procedure
that is sound, complete, and decidable.
26Systems with Changing DimensionMovie of Pursuit
Evasion
- Vehicle simulated on laptops over wireless LAN
27Scalable Information ManagementDistributed
Estimation over the GDMN
Theorem 2 If observations stop coming in at time
m and reach all the map-builders by time n,
then If communication between servers is faster
than the target then the Distributed computations
converge to the centralized ones.
Movie of our Implementation 4 servers on 4
laptops over wireless
28Summary
- Our work in the AINS program has attacked the
problem of collaborative vehicle control from two
ends - We first attacked the large-scale problem in
simulation - Then the small scale problem with real vehicles
- Future
- Attack the large scale multi-vehicle
collaborative control problem with real vehicles
29Large Numbers of Clients Generating Tasks
30 Research Objectives
- Real time 3D mapping of complex urban
environments - Coordinated team control
- Mapping, Surveillance, and Tracking Missions
31Performance Requirements
- Map the 3-D region
- in minimum time
- to a specified precision, with information
completeness, and safety of UAV - Information completeness
- Acquire all the information that could possibly
be obtained - Safety Fly with controlled risk of damage or
destruction
32Technical Challenges..
- Mapping in the face of imprecise sensor location
information - 3D map construction based on
- Multiple images from single camera
- Multiple images from multiple cameras
- Path Planning with incomplete maps in a 3D
environment - Real time Map should evolve fast enough for
control, control should evolve fast enough for
mapping - Should ensure information completeness
- UAV control for safe path execution in the face
of non-holonomy
33UAVs and Machine Vision
- Zu Kim
- Computer Science Division / California PATH
- University of California, Berkeley
34Machine Vision Challenges
- Computation
- Input 320x240 pixels (minimum), 1030 frames per
second - Cannot apply complex algorithms to real-time
applications - Human vision uses massive parallel computation
- Moving Camera
- All the pixels move
- Not many real-time algorithms dealing with moving
cameras - Limited Bandwidth
35Machine Vision Research with UAVs
- Mapping (mostly off-line)
- Aerial map generation (2-D and 3-D)
- Man-made feature detection and description
- 3-D detailed map generation (from ground-view)
- Camera calibration / accurate ego-motion
estimation - Target Detection / Recognition / Tracking
- Navigation (on-board processing)
- Stereo-based range map (depth map) generation
- Real-time ego-motion estimation / moving object
detection
362-D/3-D Aerial Map Building
- Stereo range map building (off-line)
- Commercial s/w available (eg. SocetSet fro BAE
Systems, Inc.) - 2-D map building
- Real-time image mosaic algorithm availableKang
Medioni 99,
37Man-made ObjectDetection and Description
- Building Road Detection and Description
Price 99
Kim 2000
383-D Map Building fromGround-View Frueh Zakhor
393-D Map Building fromGround-View Frueh Zakhor
40Interactive 3-D Map Building
41Target Detection and Tracking
- Challenges
- Target shapes and orientation vary
- Illumination condition varies
- Non-uniform background
- Most algorithms are based on pixel comparison,
hence require significant computation
42Target Detection and Tracking
43Navigation Stereo-based Range Map Generation
- Needs two forward-looking cameras
- Real-time algorithms available for PCs
- Intel OpenCV library
- Low quality
- Relatively simple but massive operation
- Specialized h/w (parallel processor) can improve
the quality
44Navigation Real-time Ego-motion Estimation and
Moving Object Detection
- Challenges
- Accurate ego-motion estimation require accurate
optical flows (that require massive computation)
and multi-frame optimization - Intrinsic ambiguity between speed and distance
(require sensor fusion)
45Machine Vision Research Focus
- 3-D map generation (off-line)
- Target detection and recognition (on-line and
off-line) - Navigation
- Real-time ego-motion estimation (use multiple
cameras) - Real-time moving object detection and tracking
46Planned Approach Hierarchical Design to Beat
Complexity
- Mapping Manager
- Team control and operator interface
- Adaptive 3-D Path Planner
- Handle
- Optimality
- Information completeness
- Map Information Structure
- Object independent grid decomposition
- Path computation
- Fast approximations for Traveling Salesman
computations - Adaptive space filling
47Planned Approach Hierarchical Design to Beat
Complexity
- Safe UAV flight controller
- Handles safety
- Refines the nominal 3D path into a safe but
informationally adequate path - Continuous weave of half ellipses
- Non-holonomic constraints
- Stick model (Dubins, Boissonnat)
- Extend to aggressive maneuvers
- 3D Map Builder
- Image mosaicing
- Ground robot 3D SLAM algorithms based on EM and
Maximum likelihood for (Thrun et al)
48Past Work Threat Search Simulation in the 2D
Setting
Launcher On Complete Architecture On
49UAV Sensor and Flight ControlChallenge
- Use imaging sensor to see stationary launchers
- If UAV flies straight with sensors on at its
minimum airspeed or more it gets shot by the
launchers - Imaging and object interpretation times are large
(seconds)
Launcher On UAV Sensor On
50Past Work Threat Search Simulation in the 2D
Setting
Launcher On Complete Architecture On
51 UAV Safe Controller Design Elliptical flight
path
- Algorithm
- Notations ? - imaging box side length, RL -
range of the launcher, (z,w) major and minor
axis of the ellipse -
52Coordinated Team Control for Autonomous Mapping,
Surveillance, and Tracking Missions
- Design of distributed control protocol for
adaptive relocation of sensor UAVs for optimal
sensing task performance - Estimate the geographical distribution of demand
for sensors - Complementary task division amongst heterogeneous
sensors - Proxy design for individual task monitoring
53Schedule
- Year 1
- - Off-line mapping in a complex 3D environment
with UGV image data feeds - Year 2
- - 2-UAV image fusion for navigation and mapping
in a 3D environment - Year 3
- - Collaborative mapping with 3 UAVs with dynamic
team control and task decomposition - Year 4
- - Collaborative mapping to navigate reconfiguring
formations - Year 5
- - Integration with project demonstration to
show integrated execution - of mapping, surveillance, or tracking
missions -
54Scalable Information ManagementDistributing the
Publisher Service
- Geographic Data Management Network
User