Autonomy Architecture for a Raven Class Telescope with Space Situational Awareness Applications

Transcription

1 Autonomy Architecture for a Raven Class Telescope with Space Situational Awareness Applications 3nd US Chinese Technical Interchange, Beijing May 15-17, 2013 Ryan D. Coder, Graduate Research Assistant Marcus J. Holzinger, Assistant Professor School of Aerospace Engineering, Georgia Institute of Technology 1

2 2

3 Space Situational Awareness SSA is [Joint Publication 3-14] SSA involves characterizing, as completely as necessary, Resident Space Objects (RSOs) Needs are articulated by National Space Policy (2010) DoD National Security Space Policy (2011) Helps to ensure [Joint Publication 3-14] Space flight safety Protecting economic interests Protecting space capabilities Protecting military operations and national interests Implementing international treaties and agreements 3

4 N umber of Objects The now-ubiquitous and interconnected nature of space capabilities and the world s growing dependence on them mean that irresponsible acts in space can have damaging consequences for all of us. Why SSA is Hard National Space Policy Space is vital to U.S. national security and our ability to understand emerging threats, project power globally, conduct operations, support diplomatic efforts, and enable global economic viability. As more nations and non-state actors recognize these benefits and seek their own space or counterspace capabilities, we are faced with new opportunities and new challenges in the space domain. Data deprived [Sabol et al & Nielsen et al. 2012] The current and future strategic environment is driven by three trends space is becoming increasingly congested, contested, and competitive. SSN sensors not centrally controlled [Nielsen et al. 2012] Space is increasingly congested. Growing global space activity and testing of China s destructive anti-satellite (ASAT) system have increased congestion in important areas in space. DoD tracks approximately 22,000 man-made objects in orbit, of which 1,100 are active satellites (see Figure 1). There may be as many as hundreds of thousands of additional pieces of debris that are too small to track with current sensors. Yet these smaller pieces of debris can damage satellites in orbit. Increased # of data product customers [Nielsen et al. 2012] Air Force analyst staffing issues [Weeden 2012] Satellite Catalog Growth UNCLASSIFIED Total: 1, Total: 4, Total: 6, Total: 9, Total: 22,000 Clear Cavalier Thule Fylingdales Millstone/ Haystack/Auxiliary Cape Cod Beale AF Space Surveillance System (AFSSS) Moron Optical Socorro GEODSS Eglin Surveillance System (MOSS) Maui Space Surveillance System (MSSS) Maui GEODSS Ascension Globus II Diego Garcia Ground-Based Electro-optical Deep Space Surveillance (GEODSS) Cobra Dane Reagan Test Site (RTS) Total Debris Uncataloged* Payloads Rocket bodies Iridium-COSMOS Collision COSMOS 2421 Breakup Chinese ASAT Test Shemya Radar to full-power ops Category of Sensor Near Earth (NE) Deep Space (DS) 5000 Dedicated ~ 25 % ~ 90 % Collateral ~ 70 % ~ 5 % Contributing ~ 5 % ~ 5 % UNCLASSIFIED 0 Figure 1. Source: Joint Space Operations Center * Uncataloged= unknown object and/or unknown origin Source: JSpOC, DoD 2011 National Security Space Policy 4

5 Need for Autonomy in SSA Dull, repetitive tasks: Modern systems make hundreds of observations nightly [Sabol et al. 2002] Developing observation schedule complex Fast timescales: Objects cross telescope field of view in seconds [Shell 2010] Dynamic local environment motivates near real time local schedule repair (e.g., weather) 5

6 Overview Motivation Telescopes Autonomy Proposed Architecture Example 6

7 Raven-class Telescope Overview Started as AFRL R&D effort Combination of COTS hardware and software Many Ravens currently in operation 1 Raven at Maui Space Surveillance Site contributes to SSN (Sabol et al. 2002) 7

8 Other Autonomous Telescopes LANL RAPTOR [Verstrand et al. 2008] NASA MCAT [Mulrooney et al. 2010] 8

9 Overview Motivation Telescopes Autonomy Proposed Architecture Example 9

10 Cognition Models Two common cognition models Observe, Orient, Decide, Act (OODA) loops developed by Col. John Boyd [Boyd 1976] NASA Goddard developed Plan, Perceive, Act (PPA) loop [Truszkowski et al. 2009] OODA PPA 10

11 Control Loop as a Cognition Model Orient Reference Generation Decide Controller Act Actuator Processing Observe Real World Filter / Estimator Sensor Processing 11

12 Autonomy Architectures Intelligent Machine Design Levels NASA IMD [Truszkowski et al. 2009] 3 Layer [Alami et al. 1998] Autonomy Architectures DARPA/ISO SARA [Lewandowski et al. 2001] JPL CLARAty [Estlin et al. 2001] Reflection Planning Mission Decision Routine Executive Hardware - Reaction Functional Cyber Functional Reflective Agents have the ability to learn Routine Agents have the ability to evaluate & plan Reactive Agents interface with hardware (e.g., control loops) 12

13 Machine Learning Background Categorized by type of feedback available [Russell and Norvig 2009]: Supervised Learns function to map input-output pairs Reinforcement Agent rewarded or punished for actions taken Unsupervised No explicit feedback provided 13

14 Constraint Satisfaction Problems Cast dynamic scheduling problem as CSP [Russell and Norvig 2009]: Solved using general purpose heuristics Partial sets that violate constraints removed Utility function used to select best alternative Used extensively in space applications: Hubble [Johnston 1990] Chandra [Brissenden 2001] Spitzer [Tyler et al. 2008] EO-1 [Sherwood et al. 2007] Identify opportunistic science Prioritize data downlinks 14

15 Overview Motivation Telescopes Autonomy Proposed Architecture Example 15

16 Distributed Sensor Networks Centralized superior when minimizing overall catalog covariance [Hobson et al., 2011] Centralized Mission Planner Current Space Surveillance Network [Hill et al., 2010] Uses knowledge of covariance Limited sensor knowledge Static Task Generation Decentralized superior to current SSN [Jayaweera et al., 2011] Distributed Mission Planner Sensor Sensor Dynamic Scheduler Dynamic Scheduler Distributed Mission Planner Distributed Mission Planner Limited covariance knowledge Excellent sensor knowledge Centralized Decentralized Robustness & Complexity 16

17 Proposed Autonomy Architecture Reflection Increasing Machine Intelligence Commanded Objectives Agent Agent Agent Central Planning Agent Space Object Catalog Routine Agent Agent Agent Other networked sensors Reaction Agent Agent Agent Raven Class Telescope 17

18 Overview Motivation Telescopes Autonomy Proposed Architecture Example 18

19 Space Object Detection To detect a Space Object, need an SNR ~ 6 Atmospheric transmittance Biggest factor without a model: Atmospheric Transmittance! Goal: Autonomously estimate transmittance for local azimuth and elevation over short time periods to enable local schedule repair / improvement 19

20 AllSky x480 KAI- 340 CCD F/1.4 Fujinon fisheye lens SQM-LU-DL HWHM: 10deg 20

21 Desired Information High Transmittance (SNR) Low Transmittance (SNR) 21

22 Learning with Response Surface Methodology Challenges: Lack of first-principles model for local micro-climate Computational effort Approach: Physics-based response surface equations [Kirby 2001] Catalog star observations selected intelligently using DoE [Box and Draper 1987] 22

23 Example Autonomy Architecture Reflection Increasing Machine Intelligence Commanded Plan RSM Agent Central Planning Agent Space Object Catalog Routine CSP Agent Other networked sensors Reaction Functional Agent Raven Class Telescope 23

24 Raven Example Continued Autonomous DoE to observe impactful catalog stars Empirical Function Fit (RSM) Empirical Probability of Detection All-Sky Camera Brightness Sensor SO of Interest Can then use p detect as an input to a CSP scheduler 24

25 THANK YOU 25