Robust Distributed Task Allocation for Autonomous MultiAgent Teams
Robust Distributed Task Allocation for Autonomous Multi-Agent Teams Ph. D. Candidate: Sameera Ponda Thesis Committee: Prof. Jonathan P. How, Prof. Mary L. Cummings, Prof. Devavrat Shah February 23, 2021
Motivation § Modern missions involve networked heterogeneous multi-agent teams cooperating to perform tasks § Unmanned aerial vehicles (UAVs) – target tracking, surveillance § Human operators – classify targets, monitor status § Ground vehicles – rescue operations § Key Research Questions: § How to coordinate team behavior to improve mission performance? § How to hedge against uncertainty in dynamic environments? § How to handle varying communication constraints? 2021/2/23 2
Problem Statement § Objective: Automate task allocation to improve mission performance 25 § Spatial and temporal coordination of team § Computational efficiency for real-time implementation 40 10 Problem Statement: § Maximize mission score § Satisfy constraints 25 25 § Decision variables: 30 § Team assignments, Service times § Key Technical Challenges: § § § Combinatorial decision problem (NP-hard) – computationally intractable Complex agent modeling (stochastic, nonlinear, time-varying) Constraints due to limited resources (fuel, payload, bandwidth, etc) Dynamic networks and communication requirements Robustness to uncertain and dynamic environments 2021/2/23 3
Planning Approaches § Optimal solution methods are computationally intractable for large problems 25 § Typically use efficient approximation methods [Bertsimas ’ 05] 40 10 § Most involve centralized planning [Bertsimas ’ 05] § Base station plans & distributes tasks to all agents § Requires full situational awareness § High bandwidth, slow reaction to local changes 25 25 § Motivates distributed planning [Sariel ‘ 05, Lemaire ‘ 04] § Agents make plans individually & coordinate with each other through consensus algorithms [Olfati-Saber ‘ 07] § Faster reaction to local information § Increased agent autonomy 30 § Key questions for distributed planning: § What quantities should the agents agree upon? § Information / tasks & plans / objectives / constraints § How to ensure that planning is robust to inaccurate information and models? 2021/2/23 4
Distributed Planning Centralized Problem: § Maximize mission score § Satisfy constraints § Decision variables: § Team assignments, Service times Distributed Problem: § Maximize mission score individually § Satisfy constraints § Decision variables: § Agent assignments, Service times § Main issues: Coupling & Communication § Agent score functions depend on other agents’ decisions § Joint constraints between multiple agents § Agent optimization is based on local information § Key challenge: How to design appropriate consensus protocols? [Johnson ‘ 10] § § Specify what information to communicate Create rules to process received information and modify plans Performance guarantees – is distributed problem good representation of centralized? Convergence guarantees – will algorithm converge to a feasible assignment? 2021/2/23 5
Distributed Planning – CBBA § Consensus-Based Bundle Algorithm (CBBA) [Choi, Brunet, How ‘ 09] § Iterations between 2 phases: Bidding & Consensus Phase 1: Build Bundle & Bid on Tasks (individual agents) 1 2 3 Phase 2: Consensus (all agents) All agents consistent? Yes N No § Core features of CBBA: § Sequential greedy task selection – Polynomial-time, provably good approximate solutions § Guaranteed real-time convergence even with inconsistent environment knowledge Key Contributions – extensions to CBBA framework: 1) Time-varying score functions (e. g. time-windows of validity for tasks) 2) Guaranteeing connectivity in limited communication environments 3) Robust planning for uncertain environments 2021/2/23 6
CBBA with Time-Windows e. g. monitor status, security shifts Arrival Time-critical e. g. rescue ops, target tracking Arrival Time Score Time-window Score § In realistic continuous-time missions, have time-varying task scores Peak-time e. g. rendezvous, special ops Arrival Time § Extended CBBA to continuous-time domains [ACC 2010] § Task optimization involves decisions on task assignments and task service times § Preserves convergence properties § Embedded the algorithm into dynamic planning architecture § Real-time simulation framework for dynamic missions § Experimental flight tests for UAV/UGV teams § Demonstrates real-time feasibility 2021/2/23 7
Cooperative Distributed Planning § Often have fleet-wide hard constraints on assignments § Agent assignments coupled through joint team constraints § Example: Maintaining network connectivity in dynamic environments § Often have limited communication radius, line-of-sight requirements § As agents move around environment – dynamic networks, potential disconnects 25 10 40 25 25 § Several issues: Disconnected Network 30 § Some tasks rely on continuous connectivity (e. g. streaming live video) § Cannot perform consensus, cannot deconflict plans § How to include network connectivity constraints into distributed planner? 2021/2/23 8
Example: Baseline Scenario § Motivating example – Surveillance Mission around base station § UAVs travel to tasks and stream live video back to base station § Successful task execution relies on continuous connectivity § Limited comm radius (RCOMM) 30 0 10 10 No connectivity! 15 0 2021/2/23 No connectivity! 9
Example: Network Prediction § Conservative solution – predict network connectivity violations § Drop tasks if disconnects will occur § Only execute tasks in local vicinity – conservative 10 10 30 15 2021/2/23 10
Example: Planning with Relays § Can use some agents as communication relays! § Coordinated team behavior leads to higher mission performance § Goal: Develop cooperative planning algorithms to coordinate team 30 30 10 10 Relay 15 2021/2/23 11
CBBA with Relays § CBBA with Relays [JSAC 2012, Globecom 2011, Infotech 2011, Globecom 2010] § Generate CBBA assignments § Predict network over mission duration § Repair connectivity by creating relay tasks § Key features: § Explicit consideration of dependency constraints § Predict network topology only at select missioncritical times – avoids discretizing time § Leverages information available in CBBA consensus phase § Preserves polynomial-time and convergence guarantees § CBBA with Relays improves performance § Agents accomplish higher value tasks § Guaranteed network connectivity § Demonstrated real-time applicability 2021/2/23 Real-time experiment Field experiment i. Robot Create Pelican quad
Distributed Planning Under Uncertainty § Uncertainty in planning process § Inaccurate models (simplified dynamics, parameter errors) § Fundamentally non-deterministic processes (e. g. sensor readings, stochastic dynamics) § Dynamic local information changes § Can hedge against uncertainty to improve planning Agent Schedule Target Identification Mission Late! involves several challenges § Robust planning Tasks § Optimal solutions computationally intractable – increased dimensionality of planning problem § Non-trivial coupling of distributions – analytically intractable Time § Current approaches involve many limiting assumptions Distribution for Operator Target Identification Figure from [D. Southern, Masters Thesis, 2010] § Key questions: § How to propagate uncertainty through planner to generate agent assignments? § How to distribute planning given additional complexity due to uncertainty? § How to ensure real-time performance and computational tractability? 2021/2/23 13
Distributed Planning Under Uncertainty § Chance-Constrained CBBA – Extended CBBA to incorporate risk into planning process [ACC 2012] § Model coupling using numerical approx (sampling) § Preserves polynomial-time § Probabilistic performance guarantees for given risk § Key features: § Improved CBBA to handle non-submodular score functions (e. g. stochastic scores) [CDC 2012] § Approximate distributed agent risk given mission risk using Central Limit Theorem assumption § Improved performance under uncertainty § Higher scores within allowable risk § Distributed approximation on par with centralized § Current work is exploring dynamic aspects § Dynamic risk allocation § Model learning using Nonparametric Bayesian techniques [GNC 2012] 2021/2/23 14
Conclusion § Distributed task allocation strategies for autonomous multi-agent teams § Extended CBBA algorithm to include time-varying score functions § Addressed cooperative planning in comm-limited environments using relay tasks § Presented robust risk-aware distributed extensions to deterministic planning § Acknowledgments: § § § Prof. Jonathan How for his invaluable advice and support My committee members Prof. Cummings and Prof. Shah My collaborators and colleagues at ACL, esp. Luke Johnson and Andrew Kopeikin Aero/Astro faculty and staff Graduate Aero/Astro friends! 2021/2/23 15
- Slides: 15