Scalable sWarms of Autonomous Robots and Mobile Sensors Vijay Kumar University of Pennsylvania www.grasp.upenn.edu/~kumar www.swarms.org

Motivation Future military missions will rely on large, networked groups of resource-constrained vehicles and sensors operating in dynamic, environments E Pluribus Unum, In varietate concordia Large groups will need to operate with little direct supervision Autonomy Human interaction at the group level Biology provides many models and paradigms for group behaviors Many civilian applications already, even in manufacturing industry - large numbers of embedded computers, high data rate sensors (cameras) - Main question is how to scale up from 10’s and 100’s to 1000’s. What happens when n ­ (1-10 to 10’s to 100’s)? Research problems Self aware, localize, self organize Mobility Communication Sensing Control Tasking, interfaces SWARMS

DoD Relevance Main Benefits Potential Impact SWARMS Unmanned Inexpensive Scaleable Potential Impact Adaptive communication networks for MOUT Chem/Bio search Reconnaissance and surveillance Minefield breaching Resilient Secure Unmanned – Place machines not people in harm’s way Inexpensive – Low unit cost due to simplicity of individuals Scaleable – Adding members does not require re-programming Resilient – Attrition-tolerant, fault-tolerant due to large numbers Secure – Intelligence lies in network, so capture of individual not dangerous SWARMS

Biological Models (1) Pack of wolves surrounds larger and more powerful moose. Attack vulnerable spots while moose distracted. Predator-prey model [Korf 1992] Moose: Moves to maximize its distance from nearest wolf Wolves: Each wolf moves toward the moose and away from nearest wolf The Korf model is certainly a simplification. For example, because the moose hind legs are far more dangerous than its forelegs, it would be more realistic to weight the distance by danger. It is worth noting that this model does not require any communications between the wolves. Citation: R. E. Korf. A Simple Solution to Pursuit Games. In Proceedings of Eleventh International Workshop on Distributed Artificial Intelligence, Glen Arbor, MI, pages 183-194, 1992. The wolves surround their prey using this strategy and close in on the prey. Until the wolf behind the prey closest to it can jump on the prey. The wolves in front of the prey keep its attention by menacing it, enabling one of the wolves behind the prey to jump on its back and bring it down. SWARMS

Biological Models (2) SWARMS Flocks of birds and schools of fish stay together, coordinate turns, and avoid obstacles and each other following 3 simple rules [Reynolds 1987] Termites construct mounds as tall as 5 m to store food and house brood following 2 simple rules [Kugler 1990] Flocking rules: 1/ Maintain specified minimum separation from nearest object 2/ Match velocity (magnitude and direction) to neighbors 3/ Stay close to center of flock C. Reynolds, “Flocks, Herds, and Schools: A Distributed Behavioral Model,” Computer Graphics (21)4:25-34, 1987. [The BOIDS site that this guy maintains is interesting.] Termite mound construction rules: 0/ Metabolize bodily waste, which contains pheromones 1/ Wander randomly, but prefer direction of strongest local pheromone concentration 2/ Decide stochastically whether to deposit current load. P(drop) increases with local pheromone density and with amount of current load P. Kugler, R. Shaw, K. Vincente, J. Kinsella-Shaw, “Inquiry into intentional systems I: Issues in ecological physics.” Psychological Research, 1990 These models, like the the predator-prey model, are somewhat oversimplified. SWARMS

Biological Models (3) Honey bees and ants scouting for nests Information gathering Three simple rules Explore Rate nests Recruit Tandem run; or Transport Scalable Anonymity Decentralized Simple Evaluation Deliberation Consensus building Franks et al, Trans. Royal Society, 2002 SWARMS

DoD Relevance and History Scythians vs. Macedonians, Central Asian Campaign, 329-327 B.C. Parthians vs. Romans, Battle of Carrhae, 53 B.C. Seljuk Turks vs. Byzantines, Battle of Manzikert, 1071 Turks vs. Crusaders, Battle of Dorylaeum, 1097 Mongols vs. Eastern Europeans, Battle of Liegnitz, 1241 Woodland Indians vs. US Army, St. Clair’s Defeat, 1791 Napoleonic Corps vs. Austrians, Ulm Campaign, 1805 Boers vs. British, Battle of Majuba Hill, 1881 German U-boars versus British convoys, Battle of the Atlantic, 1939-1945 Autonomous or semi-autonomous units engaging in convergent assault on a common target Amorphous but coordinated way to strike from all directions – “sustainable pulsing” of force or fire Many small, dispersed, internetted maneuver units Stand-off and close-in capabilities Attacks designed to disrupt cohesion of adversary Swarming and the Future of Conflict, RAND, 2000 SWARMS

History (continued) SWARMS Somali insurgents vs. US commandos, Battle of the Black Sea, 1993 British swarming fire harried invasion fleet Spanish Armada in 1588 German U-boat wolfpack attacks that converged on convoys in WWII Battle of the Atlantic Swarming Soviet anti-tank networks played significant role in defeating the German blitzkrieg in the Battle of Kursk SWARMS

The SWARMS Team Ali Jadbabaie, Daniel E. Koditchek, Vijay Kumar (PI), and George Pappas A. Stephen Morse David Skelly GRASP Laboratory University of Pennsylvania Center for Systems Science Yale University Francesco Bullo Daniela Rus University of California Santa Barbara CSAIL, Massachusetts Institute of Technology S. Shankar Sastry CITRIS, University of California Berkeley SWARMS

Previous Work Talk about our collective work The limitations Why they provide logical starting points for new work Outline 1. MARS Demo 2. Acclimate (Shankar?) 3. EMBER (Daniela) 4. Francesco, Steve’s work SWARMS

Fort Benning Demonstration of Networked Robots McKenna MOUT Site December 1, 2004 Research supported by DARPA, ARO (Acclimate), ONR Joint demonstration with Georgia Tech, USC, BBN, and Mobile Intelligence

Objective Network-centric force of heterogeneous platforms Provide situational awareness for remotely-located war fighters in a wide range of conditions Adapt to variations in communication performance Integrate heterogeneous air-ground assets in support of continuous operations in urban environments SWARMS

SWARMS

McKenna MOUT Site SWARMS

Main Accomplishments But… SWARMS Single operator tasking a heterogeneous team of robots for persistent surveillance Network-centric approach to situational awareness Independent of who is where, and who sees what Fault tolerant Decentralized control But… Robots are identified Control involves maintaining “proximity graph” Sharing of information SWARMS

Cooperative search, identification, and localization Grocholsky, et al, 2004 ARO, ACCLIMATE Project SWARMS

Information Model Approximate model SWARMS

Control SWARMS

Confidence Ellipsoids + = SWARMS

UGV Trajectory SWARMS

Decentralized control, but shared information… SWARMS

ACCLIMATE SWARMS

EMBER SWARMS

Steve SWARMS

Francesco SWARMS

Taxonomy of Approaches Centralized Decentralized Vehicles identified 1. Guarantees on performance 2. Optimality Identical vehicles Scalable Anonymity, Robustness Our Goal SWARMS

Scaling up to a Swarm Paradigm

Three Overarching Themes Decentralized Anonymity Simple individuals, but versatile group SWARMS

SWARMS Objective Create a research community of biologists, computer scientists, control theorists, and roboticists Systems-theoretic framework for swarming Modeling and analysis of group behaviors observed in nature Analysis of swarm formation, stability and robustness Synthesis: Formation and navigation of artificial Swarms Sensing and communication for large, networked groups of vehicles Testbeds, demonstrations, and technology transition SWARMS

SWARMS Objective Create a research community of biologists, computer scientists, control theorists, and roboticists Systems-theoretic framework for swarming Modeling and analysis of group behaviors observed in nature Analysis of swarm formation, stability and robustness Synthesis: Formation and navigation of artificial Swarms Sensing and communication for large, networked groups of vehicles Testbeds, demonstrations, and technology transition Block Island Workshop on Cooperative Control, June 10-11, 2003 Workshop on Swarming in Natural and Engineered Systems, August 3-4, 2005 SWARMS

SWARMS Research Agenda SWARMS SWARMS SWARMS SWARMS Biology AI Robotics Organism Behaviors Swarm Architectures Vehicle Models Modeling Synthesis Analysis Theory of Swarming Multi-vehicle Sensing/Control Novel Testbeds M1, M2 A1, A2, A3 S1, S2, S3 V1, V2, V3 E1, E2, E3 T1, T2 SWARMS

SWARMS Research Agenda 1. System-Theoretic Framework (T) formal language of swarming behaviors with a grammar for composition; new formalisms and mathematical constructs for describing swarms of agents derived from the unification of methods drawn from graph theory, switched dynamical systems theory and geometry. Francesco Bullo Stephen Morse George Pappas SWARMS

SWARMS Research Agenda 2. Modeling (M) model-based catalog of biological behaviors and groups with decompositions into simple behaviors and sub groups; techniques for producing abstractions of high-dimensional systems and software tools for developing low-dimensional abstractions of observed biological group behaviors. Vijay Kumar David Skelly SWARMS

Swarming in Nature SWARMS

SWARMS Research Agenda 3. Analysis (A) stability and robustness analysis tools necessary for the analysis of swarm formation; analysis of asynchronous functioning systems and abstractions to a single synchronous process; and theory for computability and complexity for swarming facilitating the design of scalable algorithms. Francesco Bullo Ali Jadbabaie A Stephen Morse SWARMS

SWARMS Research Agenda 3. Analysis SWARMS

SWARMS Research Agenda 4. Synthesis (S) design paradigms for the specification of cost functions and coordination algorithms for high-level behaviors for navigation, clustering, splitting, merging, diffusing, covering, tracking, and evasion; distributed control algorithms with constraints on sensing, actuations and communication; and software toolkit for composition of cataloged behaviors and decomposition of synthesized behaviors with the ability to automatically infer properties of resulting behaviors. Ali Jadbabaie Dan Koditschek SWARMS

SWARMS Research Agenda 5. Sensing and communication (V) estimators for vehicle and sensor platforms to localize individual agents and groups of agents; algorithms for coordinated control in support of localization and information diffusion; and bio-inspired, sensor-based (communication-less) strategies for coordination of a swarm of vehicles. Vijay Kumar Daniela Rus Shankar Sastry SWARMS

SWARMS Research Agenda 6. Testbeds, Demonstrations and Technology Transition (E) adaptive network of micro-air vehicles for aerial surveillance of an urban environment; self-healing swarm of ground vehicles (and sensor platforms) for threat and intrusion detection; and swarms of UAVs, micro-air vehicles, and small ground vehicles for operation in urban environments. Daniel Koditschek Vijay Kumar George Pappas Daniela Rus Shankar Sastry SWARMS

Alliances ARO Institute of Collaborative Biotechnology Industry Lockheed Martin (Penn) Honeywell (Berkeley/Penn) UTRC (Berkeley) Boeing (MIT/Penn) DoD Labs AFRL, ARL, NRL SWARMS

Conclusion SWARMS will develop the basic science and technology for deploying resilient, secure teams of inexpensive, unmanned vehicles Applications Adaptive communication networks Search, reconnaissance, surveillance missions SWARMS