08-Science_of_Autonomy_and_Swarming_Challenges.pdf

ONR Science of Autonomy & Swarming Challenges

Marc Steinberg Program Officer, Science of Autonomy

Scalable Robust Collaboration- Key Research Themes

Scalable – simple individuals with minimal communications lead to versatile group Rigorous mathematical methods and tools for designing & predicting behaviors

Self-organizing structure/hierarchy appropriate to mission

Distribution A – Approved for Public Release, Distribution Unlimited

Scalable/Distributed Collaboration

Engineered – Control Theory, Decentralized Optimization, Level sets, Topology, Graphs

Biological - Swarms, Flocks/Schools, Packs

Human – Game Theory, Crowds, Decentralized Organizations

Physics – Particle flow, Statistical mechanics, Force laws

Distribution A – Approved for Public Release, Distribution Unlimited

Inspiration from Many Areas

Emergence of Leaders in Schools/Flocks/Swarms, Role of Heterogeneity

Group Hunting, Foraging, Defense by Mammals

Collective Transport

Coalition & Group Formation

Herding

Human Command & Control, Decentralized Interactions

Social Abilities in Primates, other Mammals

Need to move beyond point designs to a broad spectrum of capabilities

DoD Multi-Disciplinary Research Biology, Psychology, Oceanography, Engineering, Computer Science

-Homogenous systems -Simple behaviors -Decentralized control with local information

-Heterogeneous systems -Predator/prey relationships -Adapting Command and Control structure & hierarchy

-Robustness to dynamic environments -Time-critical problems -Human supervisory control

-Networks of humans and autonomous systems -Higher-level cognition

Participants: UPenn, USC, UC Berkeley, Georgia Tech, Stanford, CMU, MIT, ASU, Northwestern

Differing Human Interaction Roles

Operator Supervisor Customer Teammate

Swarms/DecentralizedMulti-VehicleIndividual Co-located

Typical Multi-Vehicle Supervisory Control Elements vs Swarms

Map (past, current, projected)

Timeline/Progress

Vehicle/Plan/Task Status

-Unreliable communications, State Info, projections -What info is important: Metric, topological, other visual cues?

-Individual vehicles may not know when tasks are completed -How to measure task progress in a swarm?

-Individual vehicles may no longer be important -Emergent behaviors/adaptation rather than stable central plans -Potential for human intervention to disrupt rather than help

Swarm CharacteristicsInterface Elements

Controlling Swarms- Mechanism

Group Heading Moments Formations

Area Distribution, Coverage, Diffusion

Control “Leaders”

Direct control of Macro Process Parameters

Spatial Control Beacons & “Pheromones”

Control “Predator”Herd Mediate

Mission Level Task & Forget

Open Issues

• Many “swarm” demonstrations have substantially simplified the problem

– Assume good communications, off‐board GPS/motion capture infrastructure

• What happens when you have to rely on on‐board sensing for self/swarm state?

– Either simple environments/mission tasks with larger numbers or more complex missions/environments with small numbers (<4‐20)

• Can we do both large numbers and complex missions/environments?

– We have shown we can do it once

• Is it robust, reliable, repeatable, usable, understandable?

• What is an appropriate level of interaction between humans and swarms?

– How impacted by communications/topology, swarm method, mission type, user type, etc.?

– What metaphors are most effective for humans?

– Is the human role to diagnose problems/assist the swarm (e.g., break it out of local minima), manage the swarm based on things it cannot sense/does not understand (e.g., complex tactical battlespace considerations), or just task and forget?

– How long do we trust a swarm to run on its own without intervention?

• How can controls, displays, and decision support be designed for swarm control?

– What are the most effective mechanisms for shaping/managing swarm behaviors?

– How can a human infer the intent and "situation awareness" of a swarm? How does a swarm infer the human’s intent?

– Are there any best times/modes for the operator to interact with the swarm and how can the human infer or be guided to them?

– Adversary response: how to detect when a swarm has been compromised?