Active Sensing for Terrain Classification on an Agile Robot

Transcription

1 Active Sensing for Terrain Classification on an Agile Robot Richard Voyles Collaborative Systems Lab Department of Computer Science and Engineering

2 Outline Motivation: Urban Search and Rescue TerminatorBot Robotic Platform NSF Center for Safety, Security, and Rescue Robots Terrain Classification Summary and Future Work

3 The Platform: TerminatorBot

4 Emergency Response Training Sponsored by NSF Indiana Task Force 1 (FEMA) Robotics Profs R4: Rescue Robots for Research and Response

5 Emergency Response Robots

6 Standard Tools: Core-Bored Search

7 TerminatorBot to Augment Core-Bored Search

8 Custom TerminatorBot Under Development On-board Atmel RISC Microcontroller for Local Control Teleoperated Through PDA and/or Gestures Structural Tether for Retrieval, Power, and Video/Audio/Cmd Communications Pop-Up Vision CompactFlash for Full Autonomy with on-board PDA Compact Flash Interface to PDA

9 NSF Center for Safety, Security, and Rescue Research A Collaborative Effort between the University of Minnesota and the National Science Foundation to Join with Local and National Corporations in Research and Development

10 What is an NSF I/UCRC? Cooperative Research Venture between Academe and Industry Medium-term, Industrially-Relevant Research Funding NSF Provides Administrative Money Industry Provides Seed Research Money Academe Leverages Seed Money to Attract Federal Agency Money Targeted Appropriations may Provide Block Grant Money

11 A Year in the Life of SSRRC Universities Work together or in teams to refine topics Results Companies Companies Companies Companies New topics and opportunities identified Spring Symposium Poster Session ½ Day Tutorial Field Demo & Trade Fair Standards Committee Mtg Progress Posters Proposals Proposals Proposals Submitted to I/UCRC early Fall Work begins Jan 1 Ranking by PSAC, IAB Funded Funded Projects Funded Projects Projects Fall Meeting Reports from prev. year Proposal presentations PSAC, IAB mtg. Awards made

12 Research and Development Agenda Robotic Mechanisms for Rubbled Terrain Sensors and Distributed Sensor Networks Secure, Low-Power, Wireless Networks Human Activity Monitoring Reconnaissance and Surveillance

13 Prospective Member Companies Robot Companies Sensor Companies Wireless and Networks Companies Companies with Unique Fabrication or Software Expertise Companies Established in the Homeland Security Field Companies Eager to Enter the Homeland Security Field Government Labs, Centers, Agencies, etc.

14 University of Minnesota Technical Capabilities Robot Design and Prototyping Sensor Design and Interfacing Secure Networks Human Activity Monitoring Ultra-Wideband Wireless MEMS Design and Fabrication Voyles, Papanikolopoulos, Gini, Roumeliotis, Giannakis, plus???

15 Robot Mechanisms

16 Heterogeneous Sensors Ranger Communications, control, planning, reconfiguration, navigation and scout launching Sensor data TerminatorBot CPU, GPS, sensors, actuators, communications array (1) Length: 600mm M203 grenade launcher Coordination & data routing Scout (1) Audio Sensor (2) Vibration Monitor (3) Video Reconnaissance Module (2) Magnetometer, tiltmeter Video Collaborating Heterogeneous Agents (3)

17 Chemical Plume Estimation Problem: Response to Chemical, Biological and Radiological Threats Requires Early Detection of Sources and Diffusion Patterns Answer: Estimation of Gradients from Sparsely Distributed Sensors can Predict Hotspot Locations, Suggest Re-Deployment of Sensors to Optimize Observability, and Predict Evacuation Zones Due to Airborne Drift

18 Visual Servoing and Estimation

19 Distributed Control Software Behavior priority 1 Behavior priority 1 Behavior priority 2 ARC ARC ARC 60% 20% 20% RC RC RC RC

20 Activity Recognition Based on Position and Velocity Track each pedestrian throughout the scene using Kalman filter estimates Record the position and velocity Develop a position and velocity path characteristic for each pedestrian Send a warning signal under the following conditions: Pedestrian enters a secured area Pedestrian moves in tagged way Application-specific tags Running, loitering, meeting, etc Pedestrian s motion statistics do not meet the norm of crowd behavior

21 Outline Motivation: Urban Search and Rescue NSF Center for Safety, Security, and Rescue Robots Terrain Classification with Agile Robot Summary and Future Work

22 TerminatorBot

23 TerminatorBot - Alternate Scout Two 3-DoF Arms that Stow Inside Body Dual-Use Arms for both Locomotion and Manipulation Four Locomotion Gait Classes: Swimming Gaits (dry land) Narrow Passage Gait (no wider than body) Bumpy Wheel Rolling Gait Body-Roll Dynamic Gait

24 TerminatorBot Form Factor Stowed Configuration Deployed Configuration Hemispherical side for smooth manipulation Concave claw for traction/digging

25 Visual Servoing and TerminatorBot: : Goals Visual Servoing for Navigation and Homing Fixate on (many) distant features and center (object or FOE) while moving forward (like eye-in-hand) Visual odometry (3D estimation) Visual Servoing for Object Manipulation Body-fixed camera (not eye-in-hand) Visual Servoing for Terrain Identification

26 SSD Tracking SSD(dx,dy)=Σ i, j N [ I(x+dx+i, y+dy+j)-t(x+dx+i, y+dy+j) ] 2

27 Robot s s Eye View

28 Robot s s Eye View

29 Robot s s Eye View

30 Vertical Servo Error for Different Surfaces Hard Foam Peanuts

31 Robot s s Eye View

32 Bounce Normalization 2 Features (min) 3 DOF: Homing, Terrain, Body Roll Compensate for Body Roll Assuming Fixed Features

33 Terrain Classification Approach Preprocess Classifier f : θ X f : X c

34 Feature Space Gait Bounce Normalization f : θ θ Fast Fourier Transform f : θ F Fast Fourier Transform (segments) f : θ s F s θ = [ θ 1 θ 2 θ n ]

35 Classifiers Artificial Neural Networks Backpropogation with Momentum Logarithmic Sigmoid Transfer Function Support Vector Machines Discriminant Analysis Variations in Network Structure Polynomial (x y) d Radial Basis e n where n = - x-y 2 / (2σ 2 ) Sigmoid Linear Quadratic Logarithmic tanh(κ(x y) + Θ)

36 Experimental Methods Data Collection 5 terrains (carpet, BBs, woodchips, foam, rock) gait samples collected over each terrain Feature Vector Raw (Identity Function) FFT all (Frequency Space) Classifiers Evaluation FFT segmented (Concatenated Frequency Space) Artificial Neural Networks (x2) Support Vector Machines (x3) Discriminant Analysis (x3) 50 Random Divisions of Sample Data Terrain Classification Rate = mean(mean(c(x i )) t )

37 Experimental Results Raw Classifiers Raw FFTall FFTsec ann(1) ann(2) svm(p) svm svm(s) da(l) da(q) da(g)

38 Spatiotemporal Patterns and Motion Primitives PositionArm DropBody DragBody LiftBody TouchGround PositionArm

39 HMMs As Classifiers Class 1 RBRRBYBR RYBBRYRR BBRYBBRY Train to maximize P(O λ) by adjusting λ 1 parameters A,B Class 2 YYBYRRBR BBYBBBRY RBRBYYBR Train to maximize P(O λ) by adjusting λ 2 parameters A,B? RBRYBRRB Which model best fits this observation sequence?

40 Observation Sequence f : θ F preprocess preprocess preprocess QDA classify classify classify o 1 o o 9

41 Experimental Methods Segment Classifier/ HMM Training Data X 1 QDA Classifier Trainer All Sample Data Preprocess For each Segment f : θ F HMM Training Data Testing Data X 2 X 1 X 3 QDA Classifiers O 1 O 2 HMM Trainer O 3 HMM Classifiers

42 Experimental Results Total HMM Classification Rates Carpet 0 BBs Total Classification Over Multiple Gaits Chips Carpet Foam BBs Rock Chips Foam Rock 1 Cycle 1 Cycle 3 Cycles 3 Cycles

43 General Applicability Tracked Vehicle Sample "Gait" Bounce from Tank

44 Summary and Future Work Described Search-and-Rescue Motivation Detoured to NSF Center Terrain Classification Classification nearly 90% on single gait cycle Seems generally applicable to non-legged platforms Future Work: Estimating Physical Terrain Parameters Adapting Gait to Terrain Parameters Multi-Legged, Reconfigurable Mechanisms