Model-Relative Control of Autonomous Vehicles

1
Model-Relative Control of Autonomous Vehicles

• A project of DARPAs Software-Enabled Control
  program
• John Bay, Program Manager
• Principal Investigator Richard Kieburtz
• Co-PIs Eric Wan, Antonio Baptista
• OGI School of Science Engineering,
• Oregon Health Science University
• Contracting Agency AFRL
• Contract No. F33615-98-C-3516

2
Collaborators and Subcontractors

• Collaborators
• Boeing Phantom Works, UC Berkeley, Georgia Tech,
  Honeywell
• Operational requirements of the OCP
• University of Washington/Cornell University, MIT
• State and parameter estimation
• Subcontractor MIT
• will furnish an instrumented, flight-ready model
  helicopter, enabling OGI to conduct experimental
  flight tests of SDRE and MPNC control

3
Project Objectives

• Develop environmentally-informed (EI) control
  algorithms suitable for automated aircraft
  avionics and flight control under all-weather
  conditions
• Host control algorithms on OCP middleware to
  achieve platform independence and portability
  (de-emphasized)

4
Project Status (Update) -Technical

• Environmental Scenario Atmospheric Microburst
• Incorporating realistic simulated wind fields
• Effects of wind approximations on MPNC
• Environmental Scenario Urban flight example
• High-resolution modeling
• Landing on a moving platform
• Simulated motion and trajectory optimization
• Landings on a simulated ships flight deck
• FlightLab version upgrade and ship modeling.
• Progress towards flight experiments with a small
  helicopter (X-Cell60)
• Configuration and assembly

5
Environmental Scenarios Atmospheric microburst

• Atmospheric Large-Eddy Simulation (LES) Model
• Designed for the study of small-scale atmospheric
  flows (e.g. cumulus convection, entrainment,
  turbulence)
• Calculates wind velocity fields from physical
  model and boundary conditions, non-hydrostatic,
  fully 3-dimensional, quasi-compressible
• Microburst Simulation
• Grid resolution 20 m
• Subgrid-Scale Turbulence
• Finer resolution down to 1 cm.

6
Environmental Scenarios Atmospheric microburst
Gust front horizontal velocity (vertical
cross-section)
7
Environmental Scenarios Atmospheric microburst
Flight Trajectory (75 ft/s)
Horizontal wind Velocity (ft/s)
Simulated Flight through Gust Front with MPNC
8
Environmental Scenarios Atmospheric microburst

• Test of approximations for incorporating wind
  information for MPNC training

MPNC Cost comparisons
Training scenario Vehicle speed Vehicle speed
Training scenario 60 ft/s 80 ft/s
1 (no-wind info) 49.46 178.02
2 (constant wind) 28.32 57.16
3 (resolved) 26.51 38.61
4 (actual) 24.23 48.75
SDRE 111.49 162.69
(straight flight through the gust front, wind
speed 25-65 ft/s)
9
Environmental Scenarios Urban Flight

• Large-Eddy Simulation
• Use of compressible formulation allows for
  inclusion of flow obstructions
• Grid resolution 2 m
• Applied wind accelerated to approximately 6 m/s
  in free air
• Maximum reversal velocity of approx 6 m/s
  between buildings
• Maximum absolute velocity of approx 12 m/s

10
Environmental Scenarios Urban Flight
Horizontal (E-W) velocity (horizontal
cross-section at z30m)
11
Environmental Scenarios Urban Flight
Horizontal (E-W) velocity (vertical cross-section
at y160m)
12
Environmental Scenarios Urban Flight

• SDRE, cost 46.61

13
Environmental Scenarios Urban Flight

• MPNC, cost 15.34

14
Landing on a moving platform

• Platform motion (lift and roll)
• Desired Trajectory
• Linear superposition of standard landing
  trajectory with position and attitude of platform
• Aerodynamic modeling
• Current model assumes ground forces associated
  with a horizontal (moving) platform
• MPNC training for soft landing
• Added NN inputs associated with vertical force in
  landing gear and its deviation from desired
  curve.
• Quadratic cost associated with the force
  deviation from the desired curve is minimized

15
Landing on a moving platform

• SDRE landing, cost 3858.1

16
Landing on a moving platform

• Vertical forces in landing gear and suspension
  travel

17
Landing on a moving platform

• MPNC landing, cost 1645.3

18
Landing on a moving platform

• Vertical forces in landing gear and suspension
  travel

19
Ship landing simulation in Flightlab

• Recent Flightlab enhancements

Rotor induced flow
Ship airwake
Ground vortex

• Rotorcraft/ship interaction
• ship dynamics modeling
• excludes complex wave forcing
• ship airwake model
• empirical or panel method

• Aerodynamic interaction
• wing enhanced horseshoe model
• fuselage/wing panel model
• ground panel model

20
Ship landing simulation in Flightlab

• Approaches to helicopter-ship airwake
  interactions

• A. Closely coupled model
• the ship airwake is modeled with panel method
• allows proper boundary conditions on the ship
  deck and helicopter
• does not allow flow separation ? limits
  accuracy near the ship superstructure
• approach is computationally expensive for real
  time simulation

• B. Loosely coupled model
• ship airwake is computed from an accurate
  numerical model (e.g., LES model)
• ship airwake data is then applied for
  rotor/fuselage airloads calculation (e.g., via
  table look-up or our own customized turbulence
  module)
• the effect of rotor on the ship airwake is
  neglected (a one-way aerodynamic interaction)
• approach is computationally efficient for real
  time simulation

21
Flight Experiment with X-Cell60 Platform

• MIT Subcontract to assemble and test
• Assembly currently in progress
• May 3 delivered simulator to OGI.
• July 15 expected completion and delivery to OGI

22
Project Status - Next Milestones

• Evaluate robustness of SDRE and MPNC algorithms
  by simulated flights through microburst wind
  fields (31 Mar 2002)
• Simulate landing of a helicopter under autonomous
  control on a moving deck
• Without atmospheric disturbances (30 Apr 2002)
• With modeled, turbulent wind field (15 Jun 2002)
• Assembly of X-Cell helicopter and avionics
  package
• Subcontract to MIT (31 May 2002 -gt 15 July 2002)
• Hardware-in-the-loop tests of autonomous SDRE
  control system (30 Jun 2002 -gt 15 Aug 2002)
• Flight test model helicopter to gather data for
  off-line parameter estimation (31 Jul 2002 -gt 1
  Sept 2002)
• Flight simulation with X-Cell .60 flight dynamics
  model (31 Jul 2002 lt- 1 June 2002)
• Initial Flight test maneuvers with SDRE control
  (31 Aug 2002 -gt 31 Sept 2002)
• Flight test aggressive maneuvers with SDRE
  control (30 Sep 2002 -gt 1 Nov 2002)

23
Project Plans

• Next 6 Months
• FlightLab/Ship Simulation Evaluation of closely
  and loosely coupled models
• Determine impact on control algorithms
• External specification of ship motion under
  complex wave forcing
• Empirical data for ship motion anticipated
• Simulations of Helicopter landing on Ship
• Determination of optimal trajectories and control
  archtecture.
• Modification of Control Algorithms to work with
  MIT X-Cell Helicopter Model
• X-CELL Preparation
• Assemply, Flight-Test, Paramater ID, HWIL Sim,
  (etc)

24
Project Plans

• Mid-Term flight experiments with onboard SDRE
  control of an X-Cell .60 helicopter will
  demonstrate
• Automatic control of maneuvers
• Hover in fixed position
• Recover from instability to hover.
• Takeoff, translation and landing
• Sharp 90 and 180 turns at various airspeeds
• Elliptic turn in straight line flight
• Tracking a commanded flight path

25
Project Plans (contd)

• (Contract Option)
• Final Exam demonstration
• Demonstrate model-predictive, SDRE Control of
  complex maneuvers on GaTech flight platform
  (Yamaha R-Max helicopter)
• Import R-Max flight dynamics model for use with
  OGI control design suite
• Design SDRE controller with R-Max sensors and
  actuators
• Host SDRE control software on OCP
• Simulate specified maneuvers (takeoff, path
  following, landing)
• Flight tests with the R-Max will duplicate basic
  maneuvers demonstrated in the X-Cell 60 flight
  tests
• Host MPNC control software on OCP

26
Project Plans (contd)

• Additional experiments with the X-Cell flight
  platform
• SDRE Landing on a 15 inclined ramp
• SDRE 360 roll in horizontal flight
• MPNC control of pre-planned maneuvers (offline
  training)
• Additional simulation experiments
• Robustness of control
• Introduce errors in airframe mass and other model
  parameters
• Continued evaluation of wind approximations on
  control optmization
• Integrating atmospheric modeling with a FlightLab
  helicopter model
• If necessary consider a closely coupled LES
  model for environment, ship, and helicopter
• Robustness of maneuvers in turbulent wind fields
• Landing on a ships deck in rough sea and weather
  conditions

27
Project Schedule
May 1
28
Technology Transition

• Validating the control technology
• Flight tests to evaluate SDRE and MPNC control
  technology on a high-performance rotorcraft
• Portability of the software technology
• Phased transition to a middleware base on
  OCP/Build 2
• ONR FNC UAV Autonomy program
• Contract Sigma Point Kalman Filter Based Sensor
  Integration, Estimation and System Identification
  for Enhanced UAV Situational Awareness
  Control, PIWan
• Uses SPKF (UKF) for state and model estimation.
• OGIs helicopter simulator (FlightLab) and
  control system to be used for testing prior to
  transitioning to ONRs VTAUV vehicle.
• Final Exam demonstration (Contract Option)
• Demonstrate model-relative, SDRE Control of
  complex maneuvers on GaTech flight platform
  (Yamaha R-Max helicopter)

29
Program Issues

• Probable 6-week delay of Mid-term experiment
  results until October 31, 2002 caused by late
  arrival of FY2002 funding increment and delay in
  MIT subcontract.

30
End
31
Extras
32
Wind Approximations in MPNC Training

• Approximations for incorporating wind information
  for MPNC training
• MPNC trained with no information available about
  the wind.
• MPNC trained using constant wind fields as
  measured from the start of each horizon.
• MPNC trained using resolved wind field
  (turbulence is assumed unknown and neglected for
  training purposes).
• MPNC trained using knowledge of both resolved and
  actual turbulent wind flow (ideal case).

33
Landing on a moving platform

• Desired trajectory generation (landing phase)
• Smooth desired approach generated as for landing
  on a fixed flat surface,
• Desired descent curve is generated
  taking into account platform vertical motion
  (assuming Z pointing downwards)
• where ? and ? are annealed from 0 to 1. This
  allows gradual smooth transition to the landing
  phase and approach to the moving platform.
• To provide leveled landing by matching the
  platform attitude, desired pitch and roll are
  generated as functions of the platforms pitch
  and roll and the aircraft attitude. Given
  coordinates of the normal vector to the platform
• To eliminate discontinuity in desired roll and
  pitch, they are annealed using ?

34
MPNC training

• Direct signal flow diagram

35
MPNC training

• Adjoint system

36
Environmental Scenarios Urban Flight

• SDRE, cost 46.61

37
Environmental Scenarios Urban Flight
Horizontal (E-W) velocity (horizontal
cross-section at z30m)