36
The Autonomous Control and Navigation of a Trained Canine Winard “Win” Britt Committee Co-Chairs: John A. Hamilton, Jr., Department of Computer Science David M. Bevly, Department of Mechanical Engineering Committee: Saad Biaz, Department of Computer Science Outside Reader: Paul Waggoner, Canine Detection Research Inst. 1

The Autonomous Control and Navigation of a Trained Canine

  • Upload
    ozzy

  • View
    30

  • Download
    0

Embed Size (px)

DESCRIPTION

The Autonomous Control and Navigation of a Trained Canine. Winard “Win” Britt Committee Co-Chairs: John A. Hamilton, Jr., Department of Computer Science David M. Bevly , Department of Mechanical Engineering Committee : Saad Biaz , Department of Computer Science - PowerPoint PPT Presentation

Citation preview

Page 1: The Autonomous Control and Navigation of a Trained Canine

1

The Autonomous Control and Navigation of a Trained Canine

Winard “Win” BrittCommittee Co-Chairs: John A. Hamilton, Jr., Department of

Computer ScienceDavid M. Bevly, Department of

Mechanical EngineeringCommittee: Saad Biaz, Department of Computer

ScienceOutside Reader: Paul Waggoner, Canine Detection

Research Inst.

Page 2: The Autonomous Control and Navigation of a Trained Canine

2

Acknowledgement This project is financially supported by

the Office of Naval Research YIP award N00014-06-1-0518.

My doctoral studies were partially supported by the Information Assurance Scholarship Program.

Page 3: The Autonomous Control and Navigation of a Trained Canine

3

Special Thanks The engineers of the GAVLAB The canine trainers and veterinary

professionals at the Canine Detection Research Institute

The past and present undergraduates of Team K9

Page 4: The Autonomous Control and Navigation of a Trained Canine

4

Outline Introduction & Key Contributions System Architecture

Canine Hardware & Sensors Software & Control Algorithm

Experiments, Results, and Discussion Concluding Remarks

Page 5: The Autonomous Control and Navigation of a Trained Canine

5

Problem StatementA canine trained to respond to audio and

vibration commands can be autonomously directed to given waypoints without human guidance.

Page 6: The Autonomous Control and Navigation of a Trained Canine

6

Motivation Canines have sophisticated built-in

sensors for the detection of narcotics and explosives with a high degree of accuracy and at better range than people. Numerous applications.

Most canine teams require one or more support staff per canine deployed in leash-range.

If K-9s could be made to be largely autonomous, they could be used without direct human supervision keeping canines and people safer.

Page 7: The Autonomous Control and Navigation of a Trained Canine

7

Unique Challenges Dogs do not exhibit deterministic behavior

(like vehicles and robots) and are influenced by prior training and their environment in ways that robots are not.

Sensor data from a command pack is a subset of the data a human operator has when guiding a dog.

Hardware must be small, comfortable, and be able to withstand canine abuse.

Gathering canine field trial data is slow work.

Page 8: The Autonomous Control and Navigation of a Trained Canine

8

Key Contributions Developed a system to autonomously

and remotely command a trained canine using non-invasive actuation.

Demonstrated the autonomous control system through field trials with a live canine.

Demonstrated scenarios in which autonomous control of a canine surpasses that of the human operator.

Page 9: The Autonomous Control and Navigation of a Trained Canine

9

Related Work K-9 units have been used (not autonomously)

as a means of detection of explosives and narcotics with tremendous success.

Many sensor schemes to detect and analyze the pose and movements of animals.

Machine Learning has been successfully applied to a wide variety of classification and modeling problems, including a number of efforts relating to vehicle and robotic control.

Page 10: The Autonomous Control and Navigation of a Trained Canine

10

System Architectures

Two relevant phases consist of: Remote command and navigation phase Autonomous K-9 Phase

Page 11: The Autonomous Control and Navigation of a Trained Canine

11

The Trained Canine Male Labrador Retriever, 4 years old, 32

kg Trained to perform “blind retrieves” Trained to perform explosive (C4)

detection, which takes precedence over other training.

Directional training came last.

Page 12: The Autonomous Control and Navigation of a Trained Canine

12

Sensor Pack

Command Module

Rabbit Processor

XSensGPS Receiver

Xbee Radio Modem

Data Sink

GPS Satellites

Handset for

Command Module Operator:

Starts/Stops recording data for various experimental trials.

Trainer: Issues the tone and vibration commands for “left”, “right”, “stop”, “recall”, and “forward” to guide the K-9 through his handset.

Radio: Transmits the parsed sensor information and the currently active commands over the wireless link at 38400 bps.

Canine Major: Responds to the tone to follow along the intended path.

Handset: transmits the current command wirelessly to the tone generator.

Rabbit 4100: Collects and parses the sensor data from the various sensors and command module, then sends to the Xbee modem.

UBLOX GPS: Provides latitude, longitude, velocity, and heading.

XSens: Provides filtered acceleration and orientation data.

A “Remote Controlled” Canine

Command Module: Issues tone commands to the K-9 and outputs those commands to the Rabbit.

Page 13: The Autonomous Control and Navigation of a Trained Canine

13

Remote Command and Navigation phase

Develop, test, and refine hardware and software

Demonstrate a remote-controlled K-9 unit.

Record and quantify human-directed canine trials.

Understand the limitations of the canine and to be able to estimate the success rate a human operator can garner in field trials with conditions favorable to humans.

Page 14: The Autonomous Control and Navigation of a Trained Canine

14

Sensor and Command Data Summary

Page 15: The Autonomous Control and Navigation of a Trained Canine

15

Human Guided Trial Setup Establish a series of actual and foil

waypoints. Measure location using GPS. Success is defined as the human

successfully remotely guiding the canine to each waypoint (within 7m) in succession.

A “one point” failure is not arriving at even the first waypoint. A “multi-point” failure is arriving at some, but not all waypoints.

Page 16: The Autonomous Control and Navigation of a Trained Canine

16

Sample Trial

Page 17: The Autonomous Control and Navigation of a Trained Canine

17

Human Guided Trial Results

•Difference between “simple trial” success rate and “complex trials” not statistically significant (p = 0.34).•The 2-11-09 trial set is anomalous (p = 0.006). •Overall mission failure rate is about 33%.

Page 18: The Autonomous Control and Navigation of a Trained Canine

18Sensor Pack

Command Module

Rabbit Processor

GPS Receiver Xbee Radio

Modem

Data Sink

GPS Satellites

Operator: Inputs destination coordinates. Starts/Stops recording data. Operates control algorithm.

Radio: Transmits the parsed sensor information and the currently active commands over the wireless link to the laptop. Transmits back commands from controller on laptop.

K-9: Responds to the tone to follow along the intended path.

Rabbit: Collects and parses the sensor data from the various sensors, filters that data, reads new commands from the Xbee radio, issues those commands to the tone generator, then sends data back through the Xbee modem.

GPS: Provides latitude, longitude, velocity, and heading.

Autonomous K-9

Command Module: Receives commands from the rabbit and issues them to the K-9.

Page 19: The Autonomous Control and Navigation of a Trained Canine

19

Autonomous Canine Phase Develop, test, and refine control algorithms. Perform trials to validate the effectiveness

of the approach in terms of ability to get the canine to the goal waypoints.

Different paths and environments will be used in order to validate the control algorithm.

Compare autonomous guidance to human guidance.

Page 20: The Autonomous Control and Navigation of a Trained Canine

20

Goals of the Autonomous Control Algorithm

Should always give the correct commands to the canine in a timely manner.

Should not "overload" the canine with commands

Should have sensitivity to anomalous behavior, but enough leniency to account for normal variations in canine behavior (low false-positives)

Should decouple canine performance from the skill of the canine operator

Page 21: The Autonomous Control and Navigation of a Trained Canine

21

The Autonomous Control Algorithm A state-based control

developed from analysis of the human guided trials.  "States" in this case are "behaviors".

Transitions between states are events: either changes observed in canine behavior or the completion of mission related tasks.

Events are observed by trends in GPS sensor data.

Changes in states are accompanied by a new command actuation (a new tone or vibration)

High Level Control Algorithm State Machine View

Page 22: The Autonomous Control and Navigation of a Trained Canine

22

Advantages of a State Machine The idea of states and transitions is

understandable by the trainers and veterinary experts, allowing their insight to be more readily captured in the form of rules.

Tweaking system parameters from data analysis and field trials is straightforward.

False positives (calling non-anomalous behavior anomalous) are virtually non-existent. 

Page 23: The Autonomous Control and Navigation of a Trained Canine

23

Anomaly Detection Anomaly detections are caused either by undetectable

obstacles (rare), canine error(common), or sensor error (uncommon). 

Maintain state variables based on desired heading (as calculated to the next waypoint) and distance from the next waypoint. 

Sustained increasing distance from the target and/or sustained deviation from the target will cause a stop and a new directional command to be given.

Large angular deviations (going the wrong way completely) and deviations following turns (the canine did not take the turn command) will be corrected much more quickly (sub 1 second) than normal anomalies (1.5 s).

Page 24: The Autonomous Control and Navigation of a Trained Canine

24

Anomaly Correction Tricky business - the human trainer typically "recalls" after

anomalies Typically the canine makes mistakes (wrong turns) for a

reason - he wants to search something.   Issue: If canine ignores the tones/vibrations once, how

likely is it that he will suddenly stop ignoring them.  Risk could be mitigated with additional hardware.

Page 25: The Autonomous Control and Navigation of a Trained Canine

25

Autonomous Canine Trials Under “Fair” Comparison

•Difference between “simple trial” success rate and “complex trials” not statistically significant (p = 0.51).•Difference between human and autonomous canine guidance is not statistically significant (p = 1.0)

Page 26: The Autonomous Control and Navigation of a Trained Canine

26

Analysis of Missions The control algorithm issued the correct

command in all cases but one (the “Doggie U-Turn”)

Common failure to respond to turns were caused by the canine looking in the wrong direction (could be mitigated with an additional sensor on his head)

Even in environments with buildings blocking some of the line of sight to GPS satellites, sensor performance was sufficient to complete trials.

Page 27: The Autonomous Control and Navigation of a Trained Canine

27

“Unfair” Scenarios

•Some trials are tricky (impossible?) for the human to perform without line of sight.

Page 28: The Autonomous Control and Navigation of a Trained Canine

28

Future Work Address the “Doggie U-Turn” with the

Xsens Improve radio range, bandwidth and

then demonstrate in longer range scenarios.

Demonstrate system on multiple canines.

Page 29: The Autonomous Control and Navigation of a Trained Canine

29

Conclusion

I was able to demonstrate the autonomous command of a trained canine using non-invasive methods.

Automating the guidance of a canine is a complex, cross-disciplinary task that required expertise and contributions from several fields.

Page 30: The Autonomous Control and Navigation of a Trained Canine

30

Questions Questions? Comments? Nice comments are nice

too.

Page 31: The Autonomous Control and Navigation of a Trained Canine

31

Backup Slides

Page 32: The Autonomous Control and Navigation of a Trained Canine

32

Machine Learners in One Slide Given a numeric vector of input

“features”, predict one to many desired outputs.

Output must be correlated to the features!

Two phases: Development or “training” of a model from

existing data with known answers. Application of the model on new data

where the answers are unknown.

Page 33: The Autonomous Control and Navigation of a Trained Canine

33

Sensor Pack

Binary Tone

Generator

Rabbit Processor

IMUGPS Receiver

Xbee Radio Modem

Data Sink

GPS Satellites

Handset for

Binary Tone

Generator

Operator: Records the tone changes manually.

Trainer: Issues the tone commands to guide the K-9 through his handset.

Radio: Transmits the parsed sensor information over the wireless link.

K-9: Responds to the tone to follow along the intended path.

Handset: transmits the current command wirelessly to the tone generator.

Rabbit: Collects and parses the sensor data from the various sensors, then sends to the Xbee modem.

GPS: Provides latitude, longitude, velocity, and heading.

IMU: Provides acceleration and rate of turn.

Phase I (Legacy Training)

Page 34: The Autonomous Control and Navigation of a Trained Canine

34

Results of Training Phase I Verified K-9 Training and Responses to tones. Verified that reasonable sensor data could be

obtained from GPS on-board the K-9.• Created a successful (85% accurate) model of K-9 behavior using General Regression Neural Networks and Evolutionary Computation [Britt, Bevly 2008] using only available sensor inputs.

Page 35: The Autonomous Control and Navigation of a Trained Canine

35

Rationale for Departure from Initial Neural Network Approach

Training data is noisy (human guidance is inconsistent) which leads to false positives.

Difficult to get a large, robust set of specific anomalies to model effectively.

Difficult to tune a neural network in any intuitive way, even when expert knowledge is available

Page 36: The Autonomous Control and Navigation of a Trained Canine

36

Blind Field Trial Results Small number of trials performed to guarantee that the

canine could be commanded even without any waypoints In a wide, open, featureless field. No markings (not even small artificial markings) were

present on waypoints The canine was given arbitrary initial heading Average distance from goal waypoint: 13.5 m