Idea Transcript
Intelligent Control Approaches for UAVs
K. KrishnaKumar NeuroEngineering Laboratory NASA Ames Research Center
Presented at UAV-MMNT03 K. KrishnaKumar
NeuroEngineering Lab
Presentation Outline
• Intelligent Control Background • Intelligent Flight Control Research @ NASA Ames
K. KrishnaKumar
NeuroEngineering Lab
• Intelligent Control Background – What are intelligent systems – What is intelligent control – Intelligent control architectures
K. KrishnaKumar
NeuroEngineering Lab
Defining Intelligent Systems ¾ An Intelligent System is one that exhibits any of the following traits: 9 Learning 9 Adaptability 9 Robustness across problem domains 9 Improving efficiency (over time and/or space) 9 Information compression (data to knowledge) 9 Extrapolated reasoning
IS is seen as Rationalistic AI: Intelligence for doing the right thing K. KrishnaKumar
NeuroEngineering Lab
Intelligent Control U_desired ?
Y_desired ?
System
Control U
Y
Two Error Signals are needed: 1. System Performance Error Signal 2. Control Error Signal
K. KrishnaKumar
NeuroEngineering Lab
Questions How do we say that one controller is more intelligent than the other? Can the intelligence be improved? Can intelligence be measured?
Answer : Levels of Intelligent Control
K. KrishnaKumar
NeuroEngineering Lab
Levels of Intelligent Control System Features
U +_
System or Plant
Y
Mission Planning Trajectory Optimization Flight Path Adaptive Control Flight Path Stabilization
Level 0
Y desired
Level 1 Level 2 Level 3
L ev 0
Self im provem ent of: T rackin g E rror (T E )
1
T E + C on trol Param eters (C P)
2
T E + C P + Perform an ce M easure (PM )
3
T E + C P+ PM + Plan n in g Fun ction
K. KrishnaKumar
D escription Robust Feedback con trol (E rror ten ds to zero). Robust feedback con trol with adaptive con trol param eters (error ten ds to zero for n on -n om in al operation s; feedback con trol is self im provin g). Robust, adaptive feedback con trol th at m in im izes or m axim izes a utility fun ction over tim e (error ten ds to zero an d a m easure of perform an ce is optim ized). Level 2 + th e ability to plan ah ead of tim e for un certain situation s, sim ulate, an d m odel un certain ties. NeuroEngineering Lab
Levels · Level 0: Robust stabilization • • • • •
Gain Scheduling Supervised neuro-control Fuzzy control Mimic a controller Implicit Control
K. KrishnaKumar
NeuroEngineering Lab
Levels · Level 1: Adaptive Control • • • •
K. KrishnaKumar
Learn Systems and Controller Parameters Neural adaptive Control Adaptive inverse Control Approximate Controller error signal
NeuroEngineering Lab
Levels · Level 2: Optimal Control • • • • •
Reinforcement Learning Control Allocation Dynamic programming Linear Adaptive Critics Non-linear Adaptive Critics
y(t+1)
Aircraft + Controller
yd(t)
up(t)
Reference Model
Π e(t),up(t) e(t+1), etc
e(t), etc
K. KrishnaKumar
Critic (t+1) Critic (t)
λ (t+1)
λ (t)
дU(t)/д e(t) γ(1-ωdt)
+
+
NeuroEngineering Lab
Levels · Level 3: Planning Control (More AI-like) • • • • • • •
Strategic Planning Strategic search Mission Planning HTN: hierarchical task network Production-based cognitive architectures Decision-theoretic (MMDP) Etc..
FAILURE EVENT
Systems ID Updated Flight Dynamics / Performance Boundaries LANDING SITE SEARCH
ASAC AIRPORT DATABASE
Airport List
Sorted Feasible Airport List TRAJECTORY PLANNING Trajectory
GN&C
K. KrishnaKumar
NeuroEngineering Lab
NASA Ames Intelligent Flight Control Applications
K. KrishnaKumar
NeuroEngineering Lab
Manned Aircraft Objectives Develop flight control technologies that can automatically compensate for problems or failures when they occur Develop these technologies and capabilities in a generic sense so that they can be applied to different vehicle classes Application Platforms – B 757 class aircraft – Simulation only – F-15 – In Flight test – C-17 – Flight tests in 2004
K. KrishnaKumar
NeuroEngineering Lab
Pre-Trained Neural Networks Step 1 Integrated Vehicle Modeling Env. Rapid Aircraft Modeler (RAM)
K. KrishnaKumar
Step 2 Vortex Lattice Code (VORVIE W) Mass/Inertia Estimates (Balance)
Step 3 Levenberg Marquardt neural net Optimal Pruning Algorithm
NeuroEngineering Lab
Neural Flight Control Architectures Adaptive Critic technology Generalized control reallocation Adaptive Critic
Level 2 pilot inputs
Desired Handling Qualities Reference Model
+
Optimal Control Allocation
Sensors
Controller
Direct Adaptive “on-line” Neural Network
Level 1
Indirect Adaptive & “pre-trained” Neural Network(s)
Level 0 stability & control derivative estimates
-
K. KrishnaKumar
NeuroEngineering Lab
Level 1 Adaptive Control Equations
K. KrishnaKumar
NeuroEngineering Lab
Level 1 Control Equations
K. KrishnaKumar
NeuroEngineering Lab
Level 1 Control Equations
K. KrishnaKumar
NeuroEngineering Lab
Level 2: Optimal Control Allocation • When to allocate? – Control limit violation – Rate saturation – Control failure • How to allocate? – Optimal allocation using Linear Programming • Conventional hierarchy • Best available hierarchy
K. KrishnaKumar
NeuroEngineering Lab
Example Aerodynamic Control Authority Directional Authority
aileron_left
aileron_right
rudder_upper
rudder_lower
elevator_lob
elevator_lib
elevator_rib
elevator_rob
spoiler_lib
spoiler_lmib
spoiler_lmob
spoiler_lob
spoiler_rib
spoiler_rmib
spoiler_rmob
spoiler_rob
Pitch Conrol Authority
K. KrishnaKumar
Roll Control Authority
NeuroEngineering Lab
Linear Programming Formulation Dynamic System is defined as
[X& ] = f ( X ) + [B][u] + f Let us write
[B ][u ]
BUU BUL uU B B u +∆u = LU LL L L
trim
as
BUU BULuU B B u + LU LL L
BUU BUL 0 B B ∆u LU LL L
uU = Unlimited Control Vector from Dynamic Inverse u L + ∆u L = K. KrishnaKumar
Limited Control Vector from Dynamic Inverse NeuroEngineering Lab
L P Formulation (cont’d) What we need is help from Unlimited Control
BUU ∆uU BUL ∆u L B ∆u = B ∆u LU U LL L
Let us now define a control reallocation matrix [λ ] such that
[∆uU ] = [λ ][∆u L ]
=>
Define a linear relationship
[α ][λ1 K. KrishnaKumar
λ2
BUU BUL B [λ ] = B LL LU
[α ][λ ] = [β ]
. . λm ] = [β1
β2
. . βm ] NeuroEngineering Lab
LP Formulation (Cont’d) min λi
T
(wi λi )
Subject to
[α ][λi ] ≤ [β i ]
0 ≤ λ i ≤ λ max
and
Example: 4 control inputs w11 w [W ] = 21 w31 w41 K. KrishnaKumar
w12
w13
w22
w23
w32
w33
w42
w43
w14 w24 = [w1 w34 w44
w2
w3
w4 ]
NeuroEngineering Lab
Conventional & Best Hierarchies Elevator
Left Aileron
Right Aileron
Primary
Secondary
Secondary
Left Aileron
Primary
Secondary
Tertiary
Right Aileron
Secondary
Primary
Tertiary
Rudder
Secondary
Secondary
Primary
Elevator
Conventional
[W ]
T
* 100 = 100 100
K. KrishnaKumar
1 1 100 * 1 10 1 * 10 1 1 *
Rudder
Best [W ]T
* 100 = 100 100
1 * 1 1
1 100 1 1 * 1 1 * NeuroEngineering Lab
Implementation • Primary Cost based on “surface”
min T (w u ) u • Auxiliary Cost based on “axis error”
min u
K. KrishnaKumar
T
(c e )
NeuroEngineering Lab
Level 2 Controller
Reference Model Adaptation using an Adaptive Critic Adaptive Critic
pilot inputs
K. KrishnaKumar
Desired Handling Qualities Reference Model
NeuroEngineering Lab
Adaptive Critic Adaptive critic designs have been defined as designs that attempt to approximate dynamic programming.
J (t ) =< γJ (t + 1) > + min U (t ) u X(t) X(t+1)
System Model
u(t)
+
∂J (t + 1) ∂X (t + 1) γ ∂X (t + 1) ∂u(t )
X(t+1)
Critic
+
U(t)
γ
+ γJ(t+1)+U(t)
Critic J(t)
K. KrishnaKumar
∂U (t ) ∂u(t )
J(t+1) 1.0
X(t)
X(t)
Controller
+
NeuroEngineering Lab
Level 2 Control
y(t+1)
Aircraft + Controller
yd(t)
up(t)
Reference Model
Π e(t),up(t) e(t+1), etc
Critic (t+1) e(t), etc
K. KrishnaKumar
Critic (t)
λ (t+1)
λ (t)
дU(t)/д e(t) γ(1-ωdt)
+
+
-
NeuroEngineering Lab
Results for Series of Failures
During tactical descent (failures on one side) · 23,000’: Stab frozen at trim · 20,000’: 2 Elevators frozen at 0 deg. · 17,000’: Upper rudder hard over · 15,000’: Outboard flap fails retracted · 14,000’: Aileron frozen at 0 deg. · 13,000’: Two outboard spoilers frozen at 0 deg. When engines come out of reverse: Outboard engine seizes K. KrishnaKumar
NeuroEngineering Lab
Intelligent Maneuvering of UAVs •Goals –Provide increasingly higher levels of automation, capable of responding to changing goals and objectives, while taking corrective actions in the presence of internal or external events. –Allow pilots, ground-based operators or autonomous
executives to defer the responsibilities of performing and supervising tasks, to focus on managing goals and objectives.
K. KrishnaKumar
NeuroEngineering Lab
Intelligent Maneuvering of UAVs Level 2
Level 3
Strategic Planning
Tactical Maneuvering
Trajectory Trajectory Specialists Specialists
Maneuver Maneuver Database Database
Level 1
Autopilot
Flight Controller
Vehicle
Sensors (IRS) Sensors (ADC)
Sensors (NAV) ADC- Air Data Computer IRS - Inertial Reference System NAV- Navigational System (GPS & Visual Perception)
K. KrishnaKumar
Continuous-Time Commands & Sign Discrete-Time Commands
NeuroEngineering Lab
Flight Controller
K. KrishnaKumar
NeuroEngineering Lab
Tactical Maneuvering Performs time-critical flight path operations, which includes aggressive maneuvers in the presence of unexpected obstacles.
•Inputs –Commands •Reference Targets / Trajectory •Performance Parameters –Awareness •Threat Detection (eg. TCAS, GCAS) •Vehicle Performance Models •Outputs –Maneuver Sequence •Control Law Specific Modes & Targets •Transition Criteria •Maneuver Selection Specialists
Immunized Maneuver Selection Off-line
Memory BB Known Systems, Solutions, Requirements, etc
Learning
Building Blocks
Clonal Selection
On-line Immune Features
Optimal Maneuver
Utility Measures, Performance Criteria, Constraints, etc.
Krishna Kumar
Model Predictive Take-off and Landing
–Immunized Maneuver Selection –Heuristic-Based TSP Maneuver Selection •Maneuver Database –Elements & “Canned” Sequences K. KrishnaKumar
Eric Wan
NeuroEngineering Lab
A system-level description of the Immune System Metaphor
Antigen/Threat/Problem
Memory Information available in the DNA Molecule.
Immune Network Bone Marrow (Model)
•a priori knowledge, •Shape Space •Simulation, System •Representation issues Models, etc. (Binary, etc)
K. KrishnaKumar
Negative Selection
•Self-Nonself recognition •Discrimination •partly in T-cells.
Clonal Selection
•Definition of Antigen •Antigen-Antibody strength (fitness) definition
NeuroEngineering Lab
Tactical Maneuvering Database Contains general and aircraft specific maneuvering database elements, each corresponding to associated control laws. Pre-canned maneuver sequences represent domain expertise. •Elements –Control Law •Mode & Target Definition –Aircraft Specific •Flight Envelope Validation Logic –Specifications •Closed-Loop Predictive Models (x0, …, xf ) •Resource Allocation Table (e.g. lat, lon, ped, thr/col) •Sequences –Elements •Specified Parameters / Arguments –Transition Criteria / Termination Logic •Time-Based and/or Condition-Based –Interrupts •Abort Conditions & Abort Sequence K. KrishnaKumar
Bank to Turn Element Heading Select (coord. turn) Mode: HDGSEL Target: Heading = [arg1] deg Envelope: IAS > 180 kts, |θ| < … Model: ϕo/ϕi = τ /(τ s+1), ϕ‘max = gcos(θ)sin(φmax)/vt RAT: LAT/PED
Bank & Pull to Turn Sequence Bank Left: 0 Normal Accel. (speed control) Bank Right: +90 Mode: BANKSEL Target: Bank = 90 deg, Vz = Vz0 Envelope: mach > 0.4, |α| < … Model: φo/φi = τ /(τ s+1), φ‘max = pmax Vz = Vz0 RAT: LAT/PED
NeuroEngineering Lab
Autopilot System (Example) Vertical Speed From Altitude
FPA From Airspeed
Vertical Speed From V-Path
FPA From Vertical Speed
Pitch Rate From FPA
Fixed Vertical Speed
Fixed FPA
Pitch Rate From Body-Axis Pitch
Longitudinal Stick
Thrust From Vertical Speed Thrust From Airspeed Turn Rate From Heading
Throttle
Fixed Thrust
Turn Rate
•Longitudinal Modes –Pitch, Nz, AoA, FPA; Mach, IAS, Vertical Speed; Vertical Path, Altitude •Thrust Modes –Mach, IAS, Vspd, Thrust; Vertical Path, Altitude, FPA •Lateral Modes –Bank, Roll Rate; Heading, Track; Lateral Path •Directional Modes –Sideslip, Ny, Heading K. KrishnaKumar
NeuroEngineering Lab
Results 7500
Altitude (ft)
7000 6500 6000 5500 5000 4500 1
4
x 10
x 10
-1
-Y (ft)
-2
-4000
-2000
0
X (ft)
2000
4000
6000 1.6 1.5
Altitude (ft)
0 4
1.4 1.3 1.2 1.1
Three Chained Modes
1 1 0.5 4
x 10
2 1
-0.5
-Y (ft)
K. KrishnaKumar
1.5
0 -1
0.5 0
4
x 10
X (ft)
NeuroEngineering Lab
Strategic Maneuvering Performs long-term planning that meets dynamic mission goals and objectives, within mission constraints and performance limitations.
K. KrishnaKumar
Low-Altitude Energy Management 15
10
CROSSTRACK (nm)
•Inputs –Goals •Cost Function •Mission Constraints –Awareness •External Obstacles (weather, terrain, …) •Internal Health & Performance Limitations •Outputs –Extended Flight Plan •Waypoints / Reference Trajectory •Performance Parameters •Configuration Schedules •Trajectory Specialists –Energy Management Guidance •Tear Drop, Low-Altitude, Enroute –Evolutionary Navigation
5
0
TextEnd
-5
-10
-15 -20
-15
-10
-5 0 DOWNRANGE (nm)
5
10
15
John Bull
Obstacle Avoiding Evolutionary Navigation
NeuroEngineering Lab Accurate Automation Corporation
Optimal Way Point Computation Around Obstacles Using Evolutionary Algorithms ¾The Algorithm: ¾Step 1: Determine the obstacles that are in the path of the flight ¾Step 2: Place the waypoints for the aircraft on the circumference of the obstacles ¾Step 3: Compute the path between the start and the end using the waypoints. ¾Step 4: Compute a fitness function ¾Step 5. After “n” iterations the best set of waypoints defines the navigation path.
K. KrishnaKumar
NeuroEngineering Lab
Demo
K. KrishnaKumar
NeuroEngineering Lab
Intelligent Control for BEES Exploration of Mars using Free-flyers with sensors inspired by Nature Mars Lander Mars Flyer
Controller Objectives: ¾Maintain safe distance from the Lander and ensure local stability. ¾Point in the desired attitude and follow a trajectory to enable imaging of interesting Geological Picture. ¾Optimize long-term and short-term goals, such as minimization of fuel (long-term) and avoid collision with the Lander (short-term) ¾React to changing environments by adapting the control functionality
K. KrishnaKumar
NeuroEngineering Lab
Pre-Trained Neural Networks Step 1 Integrated Vehicle Modeling Env. CAD Designs
K. KrishnaKumar
Step 2 Cartesian Euler Code (CART3D)
Step 3 Levenberg Marquardt neural net Optimal Pruning Algorithm
NeuroEngineering Lab
Level 2 Architecture
Accn Commands
+ ARS Features
Tactical Maneuvering
P + I Error Controller
++
Strategic Planner
Dynamic Inversion
Learning Neural Network
Pretrained NNs
Control Allocation
Sensor Suite
K. KrishnaKumar
NeuroEngineering Lab
Concluding Remarks 9Intelligent control comes in many flavors 9Levels of Intelligent Control is one way of quantifying the roles of Intelligent control 9Intelligent control architectures allow for fast prototyping 9Intelligent control architectures can guarantee inner-loop stability 9For UAV application, intelligent control provides a robust way to accommodate any outer-loop architecture for planning, etc.
K. KrishnaKumar
NeuroEngineering Lab