for the Robocup Logistics League Jos Carlos Gonzlez ngel - PowerPoint PPT Presentation

Multi-Layered Multi-Robot Control Architecture for the Robocup Logistics League José Carlos González Ángel García-Olaya Fernando Fernández Planning and Learning Group April 15 th , 2020 Computer Science Department

Motivation • Autonomous robotics is complex Planning and Execution Competition for  Many types of sensors and actuators Logistics in Simulation  Deliberation to behave coherently Mlaras • Deliberation is complex  Now feasible for real systems • Use cases must be defined  How to really implement them? • Architecture for coordination Logistics Simulation [Kortenkamp et al. 2008] Multi-Layered Multi-Robot Control Architecture Introduction 1 /12 for the Robocup Logistics League Mlaras architecture

Automated-planning technique PDDL (:action pickup Plan 1: pickup(a) :parameters (?ob) 2: stack(a,b) Domain (action) :precondition (and a b (clear ?ob) (on-table ?ob) Classical Initial state (arm-empty)) • Simplest :effect (and • Fastest (holding ?ob) Planner (not (clear ?ob)) Probabilistic (not (on-table ?ob)) • Explicit probabilities (not (arm-empty)))) Temporal a (:init • Explicit durations Problem (on-table a) (on-table b) • Overlapping actions b (clear a) (clear b) • Overall and at-end (arm-empty)) Final state (:goal (and (on a b)))) [Ghallab et al. 2004] Multi-Layered Multi-Robot Control Architecture Introduction 2 /12 for the Robocup Logistics League Mlaras architecture

Mlaras architecture Control interface • Multi-Layered ARchitecture for User User monitoring commands Autonomous Systems Deliberation N Level N [González 2020] Action N State N • Focused on classical AP . . . . . . State 2 Action 2  Stochastic: monitoring and replanning Deliberation 1 Level 1  Temporal: condition annotations Action 1  Hierarchical: layered deliberation State 1 • Explicit abstraction conversions Deliberation 0 Level 0 Action 0 • Declarative use case definition State 0 Reactor-Hardware Introduction Multi-Layered Multi-Robot Control Architecture Mlaras architecture 3 /12 for the Robocup Logistics League Logistics competition

Mlaras modules • Deliberation layer overview Action H State H Stop Idle Goal Monitoring Converter Domain Domain Action H State M State M Action M Goal + Goal + Action M + Compat. Goal + Opp./Fail. Executive Search State L Action M State M + Action L State State H Action Converter Converter State L Action L Idle Stop Introduction Multi-Layered Multi-Robot Control Architecture Mlaras architecture 4 /12 for the Robocup Logistics League Logistics competition

Mlaras modules • Declarative configuration External Search Goal Action H State H solver relations Stop Idle Goal Model Monitoring Converter state Domain Agent Executive state Problem template Action decompositions State Action Converter Converter Durative conditions State L Action L State Idle Stop generalizations Introduction Multi-Layered Multi-Robot Control Architecture Mlaras architecture 5 /12 for the Robocup Logistics League Logistics competition

RoboCup logistics simulation league • Goal  Product building orders  More rings for more score • Robots  2 teams of 3 robots [Niemueller et al. 2015]  Should cooperate • Products  Base (3 colors)  From 0 to 3 rings (4 colors)  Cap (2 colors) Mlaras architecture Multi-Layered Multi-Robot Control Architecture Logistics competition 6 /12 for the Robocup Logistics League Mlaras instance

RoboCup logistics simulation league • Stations  Base, ring, cap, delivery  Errors break them  Require bases to work • Orders  Supplied by the referee  Product, gate and time window  Appear randomly  Completion gives score  Only 15 minutes Mlaras architecture Multi-Layered Multi-Robot Control Architecture Logistics competition 7 /12 for the Robocup Logistics League Mlaras instance

Mlaras logistics for RC logistics • Parts replaced  Deliberation 1 (GO_TO R1 C-CS1 IN START_AREA)  Abstraction conversions 1 (GO_TO R2 C-BS OUT START_AREA) 1 (GO_TO R3 C-CS1 OUT START_AREA) 2 (RETRIEVE_BASE_SHELF R1 C-CS1 IN B_TR) • New PDDL domain 2 (RETRIEVE_BASE R2 C-BS OUT B_BL) 3 (FEED_CAP_STATION R1 R3 C-CS1)  Based on classical planning 3 (GO_TO R2 C-RS1 IN C-BS) 4 (RETRIEVE_BASE_SHELF R1 C-CS1 IN B_TR) ‒ Centralized plan for the 3 robots 4 (FEED_RING_STATION R2 C-RS1 IN B_BL) 4 (GO_TO R3 C-RS2 IN C-CS1) [Borrajo et al. 2019] 5 (GO_TO R2 C-BS OUT C-RS1) • Opportunities management 5 (GO_TO R1 C-CS2 IN C-CS1) 5 (FEED_RING_STATION R3 C-RS2 IN B_TR) . . .  Rescheduling for new orders Logistics competition Multi-Layered Multi-Robot Control Architecture Mlaras instance 8 /12 for the Robocup Logistics League Conclusions

Mlaras logistics layer overview Referee Orders Monitoring Order schedule Level 2 Action 2 State 2 Robot actions Level 1 Action 1 State 1 Reactor Level 0 Action 0 State 0 Robots Logistics competition Multi-Layered Multi-Robot Control Architecture Mlaras instance 9 /12 for the Robocup Logistics League Conclusions

Results • Planning  Planning time of 10 seconds is enough  The three agents execute orders in parallel  In spite of using classical planning approach • Scores  Over 170 points if there are no obstructions ‒ Last winner 139 points, runner-up 32 points • Full declarative configuration Logistics competition Multi-Layered Multi-Robot Control Architecture Mlaras instance 10 /12 for the Robocup Logistics League Conclusions

Conclusions • Mlaras control architecture  Eases the development of autonomous systems  Declarative languages to ease use case refining • Planning and Execution for Logistics in Simulation  Classical planning approach  Stochastic, temporal, hierarchical and multiagent aspects  Competitive results • Future work  Better multiagent parallelization  Improve the high-level scheduler Mlaras instance Multi-Layered Multi-Robot Control Architecture Conclusions 11 /12 for the Robocup Logistics League

References • [Borrajo et al. 2019] Borrajo, D. and Fernández, S. (2019). Efficient approaches for multi-agent planning. Knowledge and Information Systems, 58(2):425 – 479 • [Ghallab et al. 2004] Ghallab, M., Nau, D., and Traverso, P. (2004). Automated planning: theory & practice. Elsevier • [González 2020] González, J. C. (2020). Multi-Layered Architectures for Autonomous Systems. Universidad Carlos III de Madrid (UC3M), Madrid, Spain • [Kortenkamp et al. 2008] Kortenkamp, D. and Simmons, R. (2008). Springer Handbook of Robotics - Chapter 8, pages 187 – 206. Springer Berlin Heidelberg • [Niemueller et al. 2015] Niemueller, T., Lakemeyer, G., and Ferrein, A. (2015). The RoboCup logistics league as a benchmark for planning in robotics. In Workshop on Planning and Robotics (PlanRob), ICAPS Mlaras instance Multi-Layered Multi-Robot Control Architecture Conclusions 12 /12 for the Robocup Logistics League

Multi-Layered Multi-Robot Control Architecture for the Robocup Logistics League José Carlos González Ángel García-Olaya Fernando Fernández Planning and Learning Group Thank you for your attention April 15 th , 2020 Computer Science Department

Problems of SoA architectures • Lack of guidelines and standards • Difficult to reuse previous works • Use cases are hardcoded by developers  However, they must be defined by end users • Hardcoded abstraction conversions • Complex and slow deliberation models • Lack of multilayer deliberation support All them slow down the development and advancement of autonomous systems Multi-Layered Multi-Robot Control Architecture for the Robocup Logistics League

Control strategies • Procedural control Action decomposition ∅  Behavioral trees  Tree sets can model use cases ? Selector  Action decompositions save deliberation time • Deliberation → . . . . . . . . . . . .  Decomposed deliberation Sequential  Higher layers: deliberative → →  Lower layers: reactive → →  Several simpler problems are easier Parallel Multi-Layered Multi-Robot Control Architecture for the Robocup Logistics League

Abstraction layers • High-level States LTH H  High states to deliberate with (PDDL)  High actions to define behaviors Actions • Low-level HTL H  Low states with data from the sensors  Low actions are instructions for the robot Deliberation LTH HTL Low state High state High actions Low actions Image raw data Door closed Approach Open door Depth raw data Handle at right Move arm . . . . . . Grip handle Turn hand Push door Multi-Layered Multi-Robot Control Architecture for the Robocup Logistics League

Managing temporal aspects Time 0 5 10 15 20 25 Speech Person in area Interruption! • Actions have duration  Can be unknown  Interruption in the middle of their execution • Overall and finish-when annotations  Similar to temporal planning without overlapping actions  Controlled by Mlaras, not by the planner Multi-Layered Multi-Robot Control Architecture for the Robocup Logistics League