From: AAAI Technical Report SS-99-05. Compilation copyright © 1999, AAAI (www.aaai.org). All rights reserved. Algorithms for the Design of Networksof Unmanned Aerial Vehicles Shankar S. Sastry University of California, Berkeley CA94720 At Berkeley we have been interested in design schemes for network of complex networks of semi-autonomousagents. These networks are characterized by interaction between discrete decision making and continuous control. The control of such systems is often organized in hierarchical fashion to obtain a logarithmic decrease in complexity associated with the design, Wehave used as examples three classes of systems to motivate the design approach: 1. Intelligent Vehicle HighwaySystems (IVHS) 2. Air Traffic ManagementSystems (ATMS) 3. UnmannedAerial Vehicles Over the last five years or so, a group of us have developed a set of design approaches which are aimed at designing control schemes which are live, deadlock free, and "safe". Our design methodologyis to be considered an alternative to the verification based approaches to hybrid control systems design, and is an interesting blend of gametheoretic ideas, fault handling in a probabilistic framework, mathematical and temporal logic and planning ideas from robotics. In this talk, I will focus on design problems involved in coordinating groups of UnmannedAerial Vehicles (UAVs). Problems to be addressed include: 1. Rapid prototyping of real time control laws: a hybrid systems design and simulation environment. 2. ModeSwitching and envelope protection. 3. Vision based control for navigation. 4. Mission planning for multi-UAVmissions. The work is joint with Claire Tomlin, John Lygeros, Datta Godbole, George Pappas, John Koo, David Shim, Joao Hespanha, Omid Shakernia, and Hyoun Jin Kim. 210 Intelligent Control Architectures for UCAVs& Probabilistic Verification AAAI Spring Symposium Palo Alto March 21, 1999 Intelligent ControlArchitectures ¯ An architecture design problem for a distributed system begins with specified safety, efficiency objectives & aims to characterize communication, sensing & control ¯ Progressto date - developed architecture for multi-UCAV coordinationandcontrol - implementation on the SHIFT simulation/verification testbed - developedcontrol laws for autonomous helicopters ¯ On-going research - multi-UCAV planning& coordinationdesign ARO DAAH04-96- 2 l.N~dl 211 Verification and Design Tools Research Topics - Design ModeVerification - Faulted ModeVerification - Probabilistic Verification ARO DAAH04-96- 3 1,N3d1 Perception and Action Hierarchies ¯ Research Topics - Hierarchical Vision - Control Aroundthe Vision Sensor - Surveillance Weare designing a perception and action hierarchy centered around the vision sensor to support the observation and control functions of air vehicles ¯ Progress to date - developedtarget recognition and tracking algorithms - developedtrajectory tracking controller using vision ¯ On going research - obstacle detection and avoidancealgorithms - formationflight using active vision ARO DAAH04-96- 212 A Design of Hierarchical Architecture for Multi-UAV Hybrid Systems Jin Kim, David Shim, Shankar Sastry Hierarchical SystemArchitecture Main Controller Mission Tactical Information for coordination Planner ~-~ ~ [ Tac!ical Planner ] ~- ARO£)AAH04-961,fl3~il 213 Application ¯ Mission Areais sweptby 2 helicopters, whichshouldvisit a sequence of targets assignedby main controller. ¯ Approach - Tactical Plannersets a series of suitable modesto ensure the current target is reached. - Trajectory Planner generates desiredpath for helicopter. - Regulatorwith suitable techniquesstabilizes the helicopter system with enhanced robustness. DesiredTrajectory ARODAAH04-961,=q3111 System Structure for each UAV set modes Trajectory Planner behavior Regulation] p - synthesis ~. ~. A,.B~ t, 1~.>.: , v..J.:, r Helicopter Dynamics rqk,O,v/,P,q, AROOAAH04-961~3~1 214 Techniquesfor Regulator Design Nonlinear Tracking Controller - nice transient and steady-state responsewith nominalcase only - very sensitive to modelchangeand disturbance - computationally expensive p- Synthesis Controller - tradeoff betweenthe robustnessand the tracking performance - satisfactory robustnessto modelperturbation - modelsthe noise characteristics of the sensor system - specifies the performance objective quantitatively ARODAAH04-96- 9 1~3dl Tool for Simulation - SHIFT ¯ object-oriented : accommodates structural design -> modularity ¯ precise interpretation as dynamicnetworksof interacting hybrid systems ¯ can serve as languagefor systemspecification and simulation, and as meansof team coordination ¯ Prototype-baseddescription type Helicopter { input.., whatwefeed to it output ... what wesee on the outside state ... whatis internal data mod, discrete o.. discrete modes of behavior export... eventlabels flow ¯ ii 1 continuous and discrete evolutic transition ... ) setup ... actions executedat create time l ARODA 215 10 Probabilistic Pursuit-Evasion GamesWith Cooperating UCAVs Jo~o Hespanha, Jin Kim, Shankar Sastry, 12 216 The "rules" of the game ARO DAAH04-96- 13 1,n3d1 The "rules" of the game Terrain: * with fixed obstacles (tree trunk like) * not accurately mapped UAVs(pursuers) capable of: * flying betweenobstacles * seeing a region around them (limited by the occlusions) Evadercapable of: * movingbetween obstacles (possibly actively avoiding detection) Objective: find the evader in minim~go~_e 1.f13d1 217 Scenarios ¯ search and rescue operations ARO DAAH04-96- 15 1,n3~1 The "rules" of the game ¯ search and rescue operations ¯ finding parts in a warehouse ARODAAH04-96- 16 1.q3,~1 218 The "rules" of the game ¯ search and rescue operations ¯ finding parts in a warehouse ¯ search and capture operations ARO DAAH04-96- 17 1.N3d1 The goal... ARODAAH04-g8- 18 219 Divide and conquer... The design of a swarmof UAVscapable of winning the game requires the solution of several problems: ¯ Developing strategies to win multi-agent pursuit-evasion games ¯ Mapbuilding by a network of agents: "Howto combinelocal mapsacquired by the individual agents into a global map?" ¯ Navigation (trajectory tracking & obstacle avoidance) "Howto movethe UAVfrom point A to point B keeping a minimumdistance from neighboring obstacles?" ¯ Sensingand actuation "Howto find the evader with the available sensors?" "Howto model and control a single UAV?" ARODAAH04-961,n3tll Hierarchical architecture A high-level planner capable of winning the gamewith certain probability... relying on individual UAVscapable of low level navigation & evader detection that rely on vision and actuation systems ARO DAAH04-961.93=11 220 2O Hierarchical architecture Hierarchical architecture A high-level planner capable of winning the gamewith certain probability... relying on individual UAVscapable of low level navigation & evader detection that rely on vision and actuation systems ARO DAAH04-961 n3~1 221 22 Strategies for pursuit-evasion games LaValle, Latombe,Guibas,et al. considereda similar problembut assumethe mapof the region is known,the pursuershave perfect sensors, andworstcase trajectories for the evaders: HowmanyUAVs are needed to win the gameinfinite 1 agentis sufficient time? 2 agents are needed (nomatterwhatstrategya single pursuer chooses,thereis a trajectoryfor the enemy thatavoidsdetect~,~t~9).,.,Ho.g6_ 23 1 ,Wtal Exploring a region to build a map Deng,Papadimitriou, et al., studythe problemof building a map(seeing all points in the region)traversingthe smallest possible distance. standard"keepwall to the right" algorithm algorithmthat takes better advantageof the cameras capabilities ARO DAAHO4-96- 24 1,fi3dl 222 A two step solution..¯ ¯ exploration followedby pursuit is not efficient ¯ sensors are imprecise ¯ worst ease assumptions the trajectories of the evadersleads to very conservative results ARO DAAH04-96- 25 1,f1381 A different approach¯¯. Use a probabilistie framework to combineexploration and pursuit-evasion games. Nondeterminism comes from: ¯ poorly mappedterrain ¯ noise and uncertainty in the sensors ¯ probabilistic modelsfor the motionof the evader and the UAVs ARODAAH04-96- 26 1,n3111 223 Problem formulation At each time t: ¯ the pursuers take measurements Yt (positions of enemies and obstacles seen) ¯ control actions ut are sent to the pursuers (motion control) A pursuit policy is a function g that maps past measurements to the next control action: 1"lt+l"= g(Yl, Y2 .... Yt) Goal: design pursuit policies that "maximize" probability of finding the evader ARODAAH04-96- 27 1,/ft=1 Optimalitycriteria ¯ Probability of finding the evader in the next T time steps (to be maximized) Jr (t0) = Prob(find an evader from o t o t o +T) ¯ Expected time needed to find the evader (to be minimized) J* = E{T* T* - time at which the evader is found ARODAAH04-96- 28 1,ft3dl 224 Optimalitycriteria ¯ Probability of finding the evader in the next T time steps (to be maximized) Io+T (J,=Pg(to) J,r(to)=l-I--IO-Pu(t) I=10 ¯ Expectedtime needed to find the evader (to be minimized) {" /-1 J* =Z tp (t)lI-[ I=l(I where ~ "~=10 (t) Prob(finding evader at time t, given that none wasfound up to t) pg ARO OAAH04-96- 29 1=q3=1 Sub-optimal policies greedypolicy =- pursuit policy that, at each instant, maximizesthe probability of finding the evaderat the next instant computationallyvery attractive because it scales well with the size of the region, numberof obstacles, etc. \1 ARODAAH04-96- 30 1,f13d1 225 Sub-optimal policies greedypolicy - pursuit policy that, at each instant, maximizesthe probability of finding the evaderat the next instant ¯ Probability of finding the evaderin finite time is one, i.e., Prob( T* < oc ) = 1 T* -= time at whichthe evader is found ¯ Expectedtime neededto find an evader is finite, e.g., J* = E{T*}< area n. of pursuers when a randomly movingevader is pursued by UAVsthat can checkany single unit of area at each instant of time ARO DAAH04-96- 31 1,fl3til Map building Problem: For given paths of all the UAVs,howto fuse the data sensed by each of them(in real-time) to build a global map? ARO DAAH04o96- 32 1,f13d1 226 Probabilistic mapbuilding Prob(maplnewdata & old data) = k Prob(newdatalmap) Prob(map[olddata) Problen~: For give~[ paths of all the UAltkrs,howto fuse the data sensed y each of themfin real-time) to b~l~I~r ~ofld~lmao? ARO DAAH04-96- 33 1=q3d1 Simulation (greedy policy) obstacles evader (positions unknown in advance) (Markov motion) Q UAVs probability of evaderat cell (see adjacent cells) (white--- low, black-= high) ARODAAH04-96- 34 1.n341 227