Naira Hovakimyan Department of Aerospace and Ocean Engineering Virginia Polytechnic Institute and State University This talk was originally given as a plenary at SIAM Conference on Control and Its Applications San Francisco, CA, June 2007 A. M. Lyapunov 1857-1918 G. Zames 1934-1997 Outline Historical Overview Limitations and the V&V Challenge of Existing Approaches Paradigm Shift Speed of Adaptation Performance and Robustness Rohrs’ Example Various aerospace applications Future Directions of Research Motivation Early 1950s – design of autopilots operating at a wide range of altitudes and speeds Fixed gain controller did not suffice for all conditions Gain scheduling for various conditions Several schemes for self-adjustment of controller parameters Sensitivity rule, MIT rule 1958, R. Kalman, self-tuning controller Optimal LQR with explicit identification of parameters 1950-1960 - flight tests X-15 (NASA, USAF, US Navy) bridge the gap between manned flight in the atmosphere and space flight Mach 4 - 6, at altitudes above 30,500 meters (100,000 feet) 199 flights beginning June 8, 1959 and ending October 24, 1968 Nov. 15, 1967, X-15A-3 First Flight Test in 1967 The Crash of the X-15A-3 (November 15, 1967) X-15A-3 on its B52 mothership X-15A-3 Crash due to stable, albeit nonrobust adaptive controller! Crash site of X-15A-3 Objectives of Feedback Stabilization and/or Tracking Robustness (measure of relative stability) Performance (comparison with desired system behavior) Historical Background Sensitivity Method, MIT Rule, Limited Stability Analysis (1960s) Whitaker, Kalman, Parks, et al. Lyapunov based, Passivity based (1970s) Morse, Narendra, Landau, et al. Global Stability proofs (1970-1980s) Morse, Narendra, Landau, Goodwin, Keisselmeier, Anderson, Astrom, et al. Robustness issues, instability (early 1980s) Egard, Ioannou, Rohrs, Athans, Anderson, Sastry, et al. Robust Adaptive Control (1980s) Ioannou, Praly, Tsakalis, Sun, Tao, Datta, Middleton,et al. Nonlinear Adaptive Control (1990s) Adaptive Backstepping, Neuro, Fuzzy Adaptive Control Krstic, Kanelakopoulos, Kokotovic, Zhang, Ioannou, Narendra, Ioannou, Lewis Polycarpou, Kosmatopoulos, Xu, Wang, Christodoulou, Rovithakis, et al. Search methods, multiple models, switching techniques (1990s) Martenson, Miller, Barmish, Morse, Narendra, Anderson, Safonov, Hespanha, et al. Landmark Achievement: Adaptive Control in Transition Air Force programs: RESTORE (X-36 unstable tailless aircraft 1997), JDAM (late 1990s, early 2000s) Demonstrated that there is no need for wind tunnel testing for determination of aerodynamic coefficients (an estimate for the wind tunnel tests is 8-10mln dollars at Boeing) Lessons Learned: limited to slowly-varying uncertainties, lack of transient characterization Fast adaptation leads to high-frequency oscillations in control signal, reduces the tolerance to time-delay in input/output channels Determination of the “best rate of adaptation” heavily relies on “expensive” Monte-Carlo runs Boeing question: How fast to adapt to be robust? Stability, Performance and Robustness Absolute Stability versus Relative Stability Linear Systems Theory Absolute stability is deduced from eigenvalues Relative stability is analyzed via Nyquist criteria: gain and phase margins Performance of input/output signals analyzed simultaneously Nonlinear Systems Theory Lack of availability of equivalent tools Absolute stability analyzed via Lyapunov’s direct method Relative stability resolved in Monte-Carlo type analysis No correlation between the time-histories of input/output signals The Theory of Fast and Robust Adaptation Main features guaranteed fast adaptation guaranteed transient performance for system’s both signals: input and output guaranteed time-delay margin uniform scaled transient response dependent on changes in initial condition value of the unknown parameter reference input Suitable for development of theoretically justified Verification & Validation tools for feedback systems One Slide Explanation of the New Paradigm System: x (t ) Am x(t ) b(u(t ) (t ) x(t )) Nominal controller in MRAC: unom (t ) (t ) x(t ) k g r (t ) Desired reference system: xref (t ) Am xref (t ) bk g r (t ) Implementable nom. cont. unomL1 (s) C(s) (t ) x(t ) k g r (t ) This is the nominal controller in the L1 adaptive control paradigm! Small-gain theorem gives sufficient condition for boundedness: xref L , if 1 Result: (1 C (s))( sI Am ) 1 b max 1 L1 fast and robust adaptation via continuous feedback! Two Equivalent Architectures of Adaptive Control x Am x bu T x A m Hurwitz, unknown T yc x System dynamics xm Am xm bk g r Ideal Controller u x k g r k g c A b * T T 1 m xm State predictor ~ ~ V eT Pe T1 V 0 error Barbalat’s lemma xˆ Am xˆ b u ˆT (t ) x yˆ cT xˆ Closed-loop state predictor with u xˆ Am xˆ bk g r yˆ c xˆ T T ˆ e Ame b x, e xm x Adaptive law: ˆ(t ) x(t ) eT (t ) Pb Bounded ym cT xm Adaptive Controller u ˆT (t ) x k g r regressor 1 e0 T ˆ e Ame b x e0 V 0 Same error dynamics independent of control choice Implementation Diagrams Reference system r xm Am xm bk g r u ˆx k g r u xm e x x Am x b(u T x) MRAC ˆ ˆ(t ) x(t )eT (t )Pbdt u cannot be low-pass filtered directly. State predictor based reparameterization Enables insertion of a low-pass filter u xˆ Am xˆ b(u ˆT (t ) x) x̂ Low-pass filter: C(s) r x Am x b(u x) T u ˆ(t ) x k g r ˆ ˆ(t ) x(t )eT (t )Pbdt x No more reference system upon filtering!!! e Stability u xˆ Am xˆ b(u ˆT (t ) x) x̂ Low-pass filter: C(s) r x x Am x b(u T x) u ˆ(t ) x k g r e ˆ ˆ(t ) x(t )eT (t )Pbdt Closed-loop sxˆ ( s) Am xˆ ( s) b (C ( s) 1){ˆT (t ) x(t )} C ( s)k g r Solving: xˆ ( s) ( sI Am ) 1 b (C ( s) 1) ˆT (t ) xˆ (t ) e(t ) C ( s)k g r Ho (s) bounded Sufficient condition for stability via small-gain theorem 1 C (s)sI Am 1 b max 1 L1 lim ( x(t ) xˆ (t )) 0 t Barbalat’s lemma Closed-loop Reference System uref (s) C(s)u* (s), u* (t ) T xref (t ) k g r (t ) Filtered ideal controller: Closed-loop system with filtered ideal controller = non-adaptive predictor: x( s ) b(C ( s ) 1) sx ( s ) Am x( s ) b uref ( s ) T x( s ) Am T x( s ) C ( s )k g r ( s ) Stability via small-gain theorem r Plant with unknown parameter kg C (s ) uref ( sI Am ) 1 b xref c y ref 1 C (s)H o s L max 1 1 C ( s) 1 Reference system Reference system of MRAC Guaranteed Transient Performance System state: Control signal: x xref u uref L L 1 2 C ( s) 1 (MRAC) lim x xref lim u uref 2 L 0 L 0 LTI System for Control Specifications r Plant with unknown parameter kg C (s ) uref 1 ( sI Am ) b r xref c y ref C (s ) uref ( sI Am ) 1 b xref c yr Design system for defining the control specs Reference system achieved via fast adaptation yref (s) c I (C (s) 1)( sI Am ) b 1 1 ydes (s) c (sI Am ) bC (s)k g r (s) T C (s ) T Virtual Plant with clean output kg T 1 (sI Am ) 1 bC (s)k g r (s) Independent of the unknown parameter Achieving Desired Specifications yref ydes uref udes L L 1 1 cT L1 k g H o ( s )C ( s ) C ( s ) T L1 L1 r k g H o ( s )C ( s ) Sufficient condition for stability 1 C ( s) H o s L max 1 1 Performance improvement min L L1 r L Guaranteed Performance Bounds Use large adaptive gain c 1 1 , u t uref t O , t 0 y t yref t O c c Design C(s) to render 1 C ( s) H o s L max sufficiently small 1 yref t ydes t O( ), uref t udes t O( ), t 0 Desired design objective 1 1 , u t udes t O( ) O , t 0 yt ydes t O( ) O c c Large adaptive gain Smaller step-size Faster CPU Time-delay Margin and Gain Margin rb (t ) 1 1 C ( s) C (s) 1 unbounded rb (t ) (t ) Point Timedelay occurs C ( s )1 T H ( s ) Plant C ( s) 1 T H ( s) H ( s) 1 C ( s) H ( s ) sI Am b T 1 b lim Tmargin Tm ( H ( s )) Lower bound for the time-delay margin Gm [1 , 2 ] Projection domain defines the gain margin Main Theorem of L1 Adaptive Controller Fast adaptation, in the presence of (1 C ( s)) H o (s) L max 1 , leads to guaranteed transient performance and guaranteed gain and time-delay margins: 1 lim u (t ) uref (t ) 0; lim x(t ) xref (t ) 0; lim Tmargin Tm ; Gmargin Gm , where Tm is the time-delay margin of C ( s) 1 T H ( s) H ( s) . 1 C ( s) Design Philosophy Adaptive gain – as large as CPU permits (fast adaptation) Fast adaptation ensures arbitrarily close tracking of the virtual reference system with bounded away from zero time-delay margin Fast adaptation leads to improved performance and improved robustness. Low-pass filter: Defines the trade-off between tracking and robustness Increase the bandwidth of the filter: The virtual reference system can approximate arbitrarily closely the ideal desired reference system Leads to reduced time-delay margin Tracking vs robustness can be analyzed analytically. The performance can be predicted a priori. Robot Arm with Time-varying Friction and Disturbance MgL cosq(t ) Iq(t ) F (t )q (t ) F1 (t )q(t ) (t ) u (t ) 2 1 Control objective : q( s) 2 r ( s) s 1.4s 1 Control design parameters 1 D( s) , k 60, 10000 s Parametric uncertainties in state-space form: 1 , 2 cos(t ) t 2 0 . 3 sin( t ) 0 . 2 cos( 2 t ) Compact set for Projection [10;10] Disturbances in three cases: 1 t sin( t ) 2 t cos( x1 (t )) 2 sin( 10t ) cos(15t ) 3 t cos( x1 (t )) 2 sin( 100t ) cos(150t ) Simulation Results without any Retuning of the Controller 1 t sin( t ) 2 t cos( x1 (t )) 2 sin( 10t ) cos(15t ) 3 t cos( x1 (t )) 2 sin( 100t ) cos(150t ) System output Control signal Verification of the Time-Delay Margin C ( s) 1 T H ( s) H ( s) 1 C ( s) With time-delay 0.02 Tm 0.0256 With time-delay 0.1 MRAC and L1 for a PI controller L1 MRAC The open-loop transfer functions in the presence of time-delay H ( s) e s s( s 1) Time-delay margin * : j * e 1 j ( j 1) ( ) * H ( j ) (e s 1 ) H ( s) C ( s) 2 s s * : Time-delay margin (e j 1) 1 2 j ( j 1) * lim * () 2 1.57 Application of nonlinear L1 theory C ( s) 1 , Tm 1.57 sec s 1 2 lim Tmargin Tm Miniature UAV in Flight: L1 filter in autopilot The Magicc II UAV is flying in 25m/h wind, which is ca. 50% of its maximum flight speed Courtesy of Randy Beard, BYU, UTAH Rohrs’ Example: Unmodeled dynamics kp 229 System with unmodeled dynamics: y ( s) u ( s) 2 s a s 30s 229 Nominal values k p 2, a 1 plant of plant parameters: Unmodeled dynamics Reference system dynamics: ym ( s ) dynamics 3 r (s) s3 Control signal: u (t ) k y (t ) y (t ) kr (t )r (t ) Adaptive laws from Rohrs’ simulations: ky (t ) k y (t )( ym (t ) y (t )), k y (0) 0.65, k *y 1 kr (t ) r r (t )( ym (t ) y (t )), k r (0) 1.14, k r* 1.5 Instability/Bursting due to Unmodeled Dynamics r (t ) 0.3 1.85 sin( 16.1t ) r(t)=0.3+1.85*sin(16.1*t) 4 3 r(t)=0.3+1.85*sin(16.1*t) 15 Kr 10 Ky 2 5 Instability -1 -2 Parameters 0 -5 -10 -3 -4 0 0 5 10 Time (sec) 15 -15 20 0 5 10 Time (sec) 15 20 Parameter drift System output r(t)=0.3+2*sin(8*t) r (t ) 0.3 2 sin( 8t ) 5 r(t)=0.3+2*sin(8*t) 4 Kr Ky 3 0 2 1 Bursting 0 -1 Parameters Output Output 1 -5 -2 -10 -3 0 5 10 Time (sec) 15 System output 0 5 10 Time (sec) 15 20 Parameters 20 Rohrs’ Explanation for Instability: Open-Loop Transfer Function Bode Plot (Rohr's Ex) System: P0 Frequency (rad/sec): 16.1 Magnitude (dB): -24.2 0 -20 System: P Frequency (rad/sec): 16.1 Magnitude (dB): -30.6 -40 -60 Magnitude (dB) Magnitude (dB) -20 P0=1/(s+1) -80 P=P0*229/(s2+30s+229) -100 -120 -60 P0=1/(s+1) -80 P=P0*229/(s2+30s+229) -100 -140 0 System: P0 Frequency (rad/sec): 16.1 Phase (deg): -86.4 -45 -90 System: P Frequency (rad/sec): 16.1 Phase (deg): -180 -135 -180 -180 System: P0 Frequency (rad/sec): 8 Phase (deg): -82.8 -45 Phase (deg) Phase (deg) System: P Frequency (rad/sec): 8 Magnitude (dB): -20.3 -40 -120 -140 0 -90 -135 -135 -225 -270 -2 10 System: P0 Bode Plot (Rohr's Ex) Frequency (rad/sec): 8 Magnitude (dB): -18.1 0 System: P Frequency (rad/sec): 8.01 Phase (deg): -139 -180 -225 -1 10 0 1 10 10 Frequency (rad/sec) 2 10 3 10 r (t ) 0.3 1.85 sin( 16.1t ) -270 -2 10 -1 10 0 1 10 10 Frequency (rad/sec) 2 10 r (t ) 0.3 2 sin( 8t ) At the frequency 16.1rad the phase reaches -180 degrees in the presence of unmodeled dynamics, reverses the sign of high-frequency gain, the loop gain grows to infinity At the frequency 8rad the phase reaches -139 degrees in the presence of unmodeled dynamics, no sign reversal for the high-frequency gain the loop gain remains bounded instability bursting 3 10 Rohrs’ Example with Projection for Destabilizing Reference Input With Projection operator(3) r(t)=0.3+1.85*sin(16.1*t) With Projection operator(3) r(t)=0.3+1.85*sin(16.1*t) 0.9 3 0.8 3 ky 3 0.6 Output 0.5 0.4 3 kr 3 0.3 0.2 0.1 2 5 10 Time (sec) 15 Ky 0 -1 -2 0 -0.1 0 Kr 1 Parameters 0.7 -3 20 0 5 10 Time (sec) r (t ) 0.3 1.85 sin( 16.1t ) 15 20 With Projection operator(1.5) r(t)=0.3+1.85*sin(16.1*t) 2.5 With Projection operator(1.5) r(t)=0.3+1.85*sin(16.1*t) 0.7 2 0.6 Output 0.5 1.5 k y 1.5 0.4 0.3 1.5 k r 1.5 0.2 0.1 0 0 5 10 Time (sec) 15 20 Parameters 1.5 1 Kr 0.5 Ky 0 -0.5 -1 -1.5 -2 -2.5 0 Improved knowledge of uncertainty reduces the amplitude of oscillations 5 10 Time (sec) 15 20 Rohrs’ Example with L1 Adaptive Controller u Design elements System 1 1 D( s) , k s 3 k r 100 || G ( s )(1 C ( s )) || L1 L 0.88 1 Tm 0.91 Predictor model ŷ y Adaptive Law r kˆy , kˆr kˆr u k y y 3r u D(s) (1 / s) u(t ) 3 (t ) Control Law Adaptive laws: ky (t ) k y (t )( yˆ (t ) y (t )), k y (0) 0.65 kr (t ) r r (t )( yˆ (t ) y (t )), k r (0) 1.14 No Instability/Bursting with L1 Adaptive Controller (L1) r1(t)=0.3+1.85*sin(16.1*t) (L1) r1(t)=0.3+1.85*sin(16.1*t) 4 0.7 3 0.6 2 0.5 1 Parameters 0.3 0.2 0 -1 kr -2 ky 0.1 r (t ) 0.3 1.85 sin( 16.1t ) 0 -3 -4 -0.1 -5 -0.2 0 5 10 Time (sec) 15 20 -6 0 5 10 Time (sec) 15 20 15 20 (L1) r2(t)=0.3+2*sin(8*t) 6 (L1) r2(t)=0.3+2*sin(8*t) 1.5 4 2 1 Parameters 0 0.5 Output Output 0.4 0 r (t ) 0.3 2 sin( 8t ) -0.5 -2 -4 kr -6 ky -8 -10 -12 -1 0 5 10 Time (sec) 15 20 -14 0 5 10 Time (sec) Difference in Open-Loop TFs ym 1 s3 Pref(s) r(t) Kˆ r (u, ~y ) + + u 229 s 2 30s 229 Pum(s) 2 s 1 + y(s) - MRAC ~y P(s) MRAC cannot alter the phase in the feedforward loop Adaptation on feedforward gain Kˆ y ( y, ~y ) + ~ ˆ K r (u, y ) + Kˆ (u, ~y ) Continuous adaptation on phase y r(t) L1 kg + + + k s Kˆ r (u , ~y ) u 229 s 2 30s 229 Pum(s) Kˆ y (u, ~y ) 2 s 1 P(s) 1 s3 Ppr (s) + y(s) - ~y Classical Control Perspective L1 MRAC Adaptation simultaneously on both loop gain and phase Adaptation only on loop gain No adaptation on phase Stabilization of Cascaded Systems with Application to a UAV Path Following Problem Ge Gp u UAV with Autopilot y y(s) G p ( s)(u( s) z (s)) u x Path Following Kinematics x f ( x) g ( x) y Gp Autopilot y UAV Cascaded system (in collaboration with Isaac Kaminer, NPS) Augmentation of an Existing Autopilot by L1 Controller Path following kinematics Autopilot Exponentially stable Path following controller Autopilot UAV UAV Path following kinematics Very poor performance Path following controller Path following kinematics Output feedback L1 augmentation L1 adaptive controller Path following controller Flight Test Results of NPS (TNT, Camp Roberts, CA, May 2007) L1 Adaptive Controller Augmented L1 Adaptive Controller u Autopilot y UAV Inner-loop Guidance and Control yc Outer-loop path-planning, navigation Without adaptation Errors 100m With adaptation, errors reduced to 5m Architecture for Coordinated Path Following • Path following: follow predefined trajectories using virtual path length • Coordination: coordinate UAV positions along respective paths using velocity to guarantee appropriate time of arrival Guaranteed stability of the complete system: use L1 to maintain individual regions of attraction Path Generation Network info desired Path following path (Outer loop) Normalized Path lengths Coordination L1 adaptation Onboard A/P + UAV (Inner loop) Pitch rate Yaw rate commands Velocity command L1 adaptation Time-critical Coordination of UAVs Subject to Spatial Constraints: Hardware-in-the-Loop Flight Test Results Performance comparison with and without adaptation for one vehicle Simultaneous arrival of two vehicles with L1 enabled Mission scenarios: Sequential autolanding, ground target suppression, etc. Arrival time is the same, but is not specified a priori Assumptions: Vehicles dynamically decoupled Underlying communication network topology is fixed in time (currently relaxed) Each vehicle communicates only with its neighbors Flight Control Design (Models provided by the Boeing Co.) X-45ATailless Unstable Aircraft Elevon/yaw vectoring control mixing Compensation for pitch break uncertainty and actuator failures Autonomous Aerial Refueling Uncertain dynamic environment Receiver dynamics due to aerodynamic influences from tanker Drogue dynamics due to aerodynamic influences from tanker and receiver Finite time to contact the drogue in highly uncertain dynamic environment. X-45A: Transient Performance vs Robustness Case 2: ROB Elevon fail @ 1 sec, Pitch Break ON • • 1 Transient tracking not too sensitive to changes in adaptive gain beyond certain lower bound Filter impacts the transient tracking The filter can be selected to improve the margin at the price of transient tracking 0 -1 -2 Az ft/sec/sec • -3 C (s ) -4 C1 ( s) C(s), =5000 C(s), =50000 C1(s), =5000 C1(s), =50000 -5 -6 -7 -8 -9 TDM=0.045 0 5 10 15 20 25 30 Time (sec) 35 40 45 Vertical acceleration Case 2: ROB Elevon fail @ 1 sec, Pitch Break ON TDM=0.1 0 • • Transient tracking compared for MRAC and L1 with two different filters The performance of C1(s) (with better margins) still better than for MRAC Az, ft/sec/sec -2 C1 ( s) -4 -6 MRAC C(s)=1/(0.05s+1) C1(s)=C(s)*(-6s+1) 2/(8s+1)2 C (s ) -8 -10 MRAC 0 5 10 15 20 25 30 35 40 45 Aerial Refueling Control objective: Finite time to contact the drogue in highly uncertain dynamic environment. Solution/Approach: An adaptive method without retuning adaptation gains for highly uncertain dynamic environment with guaranteed transient performance for finite time-to-contact • Barron Associates Tailless Aircraft Model • 6 DOF Flying-wing UAV, Statically unstable at low angle of attack • Aerodynamic data primarily taken from [1] • Vortex effect data from wind tunnel test [2] – ICE model • Tanker: KC-135R, Receiver: Tailless 65 degree delta wing UAV • Aerodynamic coefficients change as a function of relative separation → Varying more significantly with lateral and vertical separation [1] S.P Fears et. al., Low-speed wind-tunnel Investigation of the stability and control characteristics of a series of flying wings with sweep angles of 50 degree [2] W.B. Blake et. al., UAV Aerial Refueling – Wind Tunnel Results and Comparison with Analytical Predictions Scaled Response - Starting from different initial positions - Scaled response No retuning! - Twice the vortex – model of a different tanker thrust channel elevator channel aileron channel - Uniform transient and steady state response - Scaled control efforts Air-Breathing Hypersonic Vehicle (Boman model WP AFRL) Control Challenge of AirBreathing Hypersonic Vehicles Vibration due to high speed and structural flexibility Uncertainties caused by rapid change of the operating conditions Integration of airframe dynamics and propulsion system Coupling of attitude and velocity Desired path flight angle desired speed Control signal: Elevator (deg) Velocity profile System response Control Signal: Equivalent ratio Current Status and Open Problems Adaptive Output Feedback Theory Partial results in this direction Numerical discretization issues due to fast adaptation Integration of stiff systems Adaptive Observer Design Control of distributed autonomous systems over ad-hoc communication network topologies Performance evaluation on the interface of different disciplines More work on its way towards final answers!!! Conclusions What do we need to know? Boundaries of uncertainties sets the filter bandwidth CPU (hardware) sets the adaptive gain Fast adaptation no more trial-error based gain tuning! performance can be predicted a priori robustness/stability margins can be quantified analytically Theoretically justified Verification & Validation tools for feedback systems at reduced costs With very short proofs! Acknowledgments My group: Chengyu Cao (theoretical developments) Jiang Wang (aerial refueling) Vijay Patel (UCAV, flight test support) Lili Ma (vision-based guidance) Yu Lei (hypersonic vehicle) Dapeng Li (theoretical developments, vision-based guidance) Amanda Young (cooperative path following) Enric Xargay (cooperative path following) Zhiyuan Li (discretization issues for flexible aircraft) Aditya Paranjape (differential game theoretic approaches for VBG) Collaborations: The Boeing Co. (E. Lavretsky, K. Wise, UCAV model, aerial refueling) Wright Patterson AFRL, CerTA FCS program (ICE vortex) Raytheon Co. (R. Hindman, B. Ridgely) NASA LaRC (I. Gregory, SensorCraft) Wright Patterson AFRL (D. Doman, M. Bolender, hypersonic model) Randy Beard (BYU) Isaac Kaminer (NPS) Sponsors: AFOSR, AFRL, ARO, ONR, Boeing, NASA