Title: June 10, 2002
1June 10, 2002
2Adaptive Coordinated Control of Intelligent
Multi-Agent Teams
- Shankar Sastry, Ruzena Bajcsy, Laurent El Ghaoui,
Mike Jordan, Jitendra Malik, Stuart Russell,
Pravin Varaiya (Berkeley) - Vijay Kumar, Kostas Danillidis, James Ostrowski,
George Pappas, C. J. Taylor (Penn) - Howie Choset, Alfred Rizzi, Charles Thorpe (CMU)
3Background
- ARO-MURI Integrated Approach to Intelligent
Systems 1996-2001. Partners Berkeley, Stanford,
Cornell. Highlights - Creation of Field of Hybrid Systems
Foundations, Methods, Analysis, Control - Vision Based Navigation and Control
- Major Force Driving Bayesian Networks, Graphical
Models, Dynamical Probabilisitic Networks,
Learning, Rapproachment of AI and Control
4Integrated Approach to Intelligent Systems
- Disciplinary Evolution
- Control Theory
- Optimal control, linear control, nonlinear
control, adaptive control, stochastic control - Mathematics differential equations
- Artificial Intelligence
- Reasoning, adaptation, neural networks, natural
language, expert systems - Mathematics formal logic
- Computational Neuroscience Cognitive Science
- Sensing, vision, taction, olfaction neural
networks - Mathematics (recently developed)
5Why Hierarchical Hybrid Control?
- Central Control Paradigm. that is sensors and
actuators interacting locally, breaks down when
dealing with distributed systems due to - Complexity scale
- Necessity of tight or optimal operations
- Key Characteristics of Distributed Intelligent
Systems - Hierarchical or modular to control complexity
- Globally organized emergent behavior
- Robust, adaptive and fault tolerant, and degraded
modes of operation - Architectural organization involving the use of
compositionality
6Why Hybrid Hierarchical Control?
- Intelligence Augmentation for Human-Centered
Systems - Autonomous Intelligence
- Why integrative? Due to
- - the need to merge sensor fusion and
hierarchies of sensing with actuation across many
agents, with desired emergent behavior - - the need to merge logical decision making and
continuous action - - the need reconcile the need for safety of
individual agents with collective optimality - Control, artificial intelligence and cognitive
neuroscience deal with continuous action, logical
reasoning and human/machine understanding,
respectively
7Technology Drivers Semi-Autonomous Multi-Agent
Systems
- The need for a theoretical framework for an
integrative approach arises from advances in
computation, communication, intelligent
materials, visualization and other technologies
which make it possible to expect more from a
multi-agent system than from a centralized
control framework. - Distributed Command and Control
- Distributed Communication Systems
- Distributed Power Systems
- Intelligent Vehicle Highway Systems
- Air Traffic Management Systems
- Intelligent Telemedical Systems
- Intelligent Manufacturing Systems
- Unmanned Aerial Vehicle Networks
- Mobile Offshore Bases
8Theoretical Underpinnings
- Architectural Design for Multi-Agent Systems
- Hybrid Systems
- Centralization for optimality
- Decentralization for safety, reliability and
speed of response - Perception Systems Sharing Many Representations
- Hierarchical aggregation
- Wide-area surveillance
- Low-level perception
- Frameworks for Representing and Reasoning with
Uncertainty - Incorporation of Learning, Adaptation and Fault
Tolerance - Parametric uncertainty with update and adaptation
at the continuous levels, learning of new
logical entities --reinforcement learning at
the logical levels and metal-learning for
redefining architecture
9What Are Hybrid Systems?
- Dynamical systems with interacting continuous and
discrete dynamics
10Why Hybrid Systems?
- Modeling abstraction of
- Continuous systems with phased operation (e.g.
walking robots, mechanical systems with
collisions, circuits with diodes) - Continuous systems controlled by discrete inputs
(e.g. switches, valves, digital computers) - Coordinating processes (multi-agent systems)
- Important in applications
- Hardware verification/CAD, real time software
- Manufacturing, communication networks, multimedia
- Large scale, multi-agent systems
- Automated Highway Systems (AHS)
- Air Traffic Management Systems (ATM)
- Uninhabited Aerial Vehicles (UAV), Power Networks
11 Framework
12Control Challenges
- Large number of semiautonomous agents
- Coordinate to
- Make efficient use of common resource
- Achieve a common goal
- Individual agents have various modes of operation
- Agents optimize locally, coordinate to resolve
conflicts - System architecture is hierarchical and
distributed - Safety critical systems
- Challenge Develop models, analysis, and
synthesis tools for designing and verifying the
safety of multi-agent systems
13Taking Stock in Hybrid Systems
- Hybrid Systems and Control established as a
discipline, taught to undergrads, grads.
Monographs, textbooks being written by all
co-PIs Lee and Varaiya, Henzinger and Alur,
Lygeros, Tomlin and Sastry. Workshop on Hybrid
Systems established (first was in Berkeley in
1998). Special Issues in IEEE Proceedings,
Systems and Control Letters, Automatica, IEEE
Transactions on Automatic Control, - Software Programming languages, tools and
frameworks for Simulation and Control Ptolemy
II, Giotto, Massaccio all developed. Ongoing work
on verification tools. - Hardware in the loop demonstrations on the local
UAVs, formation flying to follow. - Embedded software EMSOFT established, new IEEE
Proceedings Special Issue 2003.
14UCB/UCSF Laparoscopic Telesurgical Workstation
15ANIMAL LAB TRIALS 1998
16Suturing with Unimanual System, 1998
17Berkeley BEAR Fleet Ursa Maxima 1
Based on Yamaha RMAX industrial helicopter
Integrated Nav/Comm Module
Length 3.63m Width0.72m Height 1.08m Dry
Weight 58 kg Payload 30kg Engine Output 21
hp Rotor Diameter 3.115m Flight system
operation time 60 min
18Flight Control System Experiments
Landing scenario with SAS (Dec 1999)
PositionHeading Lock (Dec 1999)
PositionHeading Lock (May 2000)
Attitude control with mu-syn (July 2000)
19Pursuit-Evasion Game Experiment Setup
Waypoint Command
Pursuer UAV
Current Position, Vehicle Stats
Evader location detected by Vision system
Ground Command Post
Current Position, Vehicle Stats
Evader UGV
20Pursuit-Evasion Game Experiment
- PEG with four UGVs
- Global-Max pursuit policy
- Simulated camera view
- (radius 7.5m with 50degree conic view)
- Pursuer0.3m/s Evader0.5m/s MAX
21Experimental Results Pursuit Evasion Games with
4UGVs (Spring 01)
22Pursuit-Evasion Game Experiment
- PEG with four UGVs and a UAV
- Global-Max pursuit policy
- Simulated camera view
- (radius 7.5m with 50degree conic view)
- Pursuer0.3m/s Evader0.5m/s MAX
23Experimental Results Pursuit Evasion Games with
4UGVs and 1 UAV (Spring01)
24What is Different Today?
- The world and national security threats are
different mobile operations in urban terrain,
hostage rescue, anti terrorism operations,
homeland protection. - Use of robotic and mixed initiative forces, the
need for coordination of manned and unmanned
forces - The need for dynamic strategies and tactics for
dealing with a determined and flexible adversary. - Exploitation of the 3rd dimension by organic UAVs
25New Technical Innovations
- Control of the 3 D Digital battlefield need to
use 3rd dimension, aerial forces, robotic and
mixed initiative forces, untethered
communications - Adaptive Coordinated Control of Multiple Agents
reconfiguration of teams dynamically in response
to adversarial action - Intelligent coordination of multiple agents
ability to discover intent and reconfigure
strategies adaptively
26Intellectual OrganizationThrust Areas
- Architecture Design for Adaptive, Dynamic
Planning - Integration of Rich Multi-Sensor Information into
Virtual Environments incorporating human
intervention - Handling Uncertainty and Adversarial Intent in
Adaptive Planning
27Challenge Scenarios
- Reconaissance and Intelligence robotic ranger
force for scouting fixed area for time critical
targets - Mixed Initiative Engagement in urban environments
using micro-UAVs, UGVs. Emphasis on immersive
environments for deploying - Recognition and Tracking of Unfriendlies
emphasis on networked vision for tracking.
28Hierarchical Architectures for Dynamic Adaptive
Planning
- Progess to date in hierarchical architectures for
decision making in normal modes of operation.
Main emphasis here will be on replanning in
fault or degraded modes of operation
including deviations from hierarchical operation. - Key technical issues
- Abstractions of Hybrid Systems for Architecture
Design - Hierarchical abstractions
- Assume-guarantee reasoning for abstractions
29Thrust I continued
- Control of Hybrid Systems
- Numerical Solutions for Controller Synthesis
- Hierarchical Solutions of Synthesis Procedures
- Liveness and other acceptance conditions
- Controller Libraries
- Many world semantics and hierarchy semantics
- Modal decomposition
- Exceptions
- Team and Task Allocation
30Integration of Multi-Sensor Information Into
Virtual Environments
- Adaptive Hierarchial Networks for Acquiring and
providing information - Networked sensors
- Bandwidth utilitzation
- Extraction of 3 D Models from Distributed Sensors
- 3 D models from video data
- Integration of real and virtual environments
- Environments for Human Intervention Decision
Making - Situational awareness
- Display of uncertain data
- Triaging of data for decision making
31Smart Dust, Dot Motes, MICA Motes
- Dot motes, MICA motes and smart dust
32Tiny OS (TOS)
- Jason Hill, Robert Szewczyk, Alec Woo, David
Culler - TinyOS
- Ad hoc networking
33MAVs for Delivery
www.spyplanes.com
34Last 2 of 6 motes are dropped from MAV
35 Field of wireless sensor nodes
- Ad hoc, rather than engineered placement
- At least two potential modes of observation
- Acoustic, magnetic, RF
36 Subset of more powerful assets
- Gateway nodes with pan-tilt camera
- Limited instantaneous field of view
37Set of objects moving through
38 Track a distinguished object
39Many interesting problems
- Targeting of the cameras so as to have objects of
interest in the field of view - Collaborate between field of nodes and platform
to perform ranging and localization to create
coordinate system - Building of a routing structures between field
nodes and higher-level resources - Targeting of high-level assets
- Sensors guide video assets in real time
- Video assets refine sensor-based estimate
- Network resources focused on region of importance
40Abstraction of Sensorwebs
- Properties of general sensor nodes are described
by - sensing range, confidence on the sensed data
- memory, computation capability, clock skew
- Communication range, bandwidth, time delay,
transmission loss - broadcasting methods (periodic or event-based)
- To apply sensor nodes for the experiments with
UAV/UGVs introduce super-nodes (or gateways),
which can - gather information from sub-nodes ( filtering or
fusion of the data from sub-nodes for partial map
building) - communicate with UAV/UGVs
41Uncertainty and Adversarial Intent
- Models of Uncertainty
- Environmental non deterministic and
probabilistic - Adversarial
- Guarantees of Success in the face of uncertainty
- Decision making in the presence of uncertainty
- Learning of Adversarial Strategy
- Probing strategies
- Games, partial information solution concepts
- Adaptation to changing utility functions of
adversary