Distributed Hierarchical Interleaved Planning and Supervision The case of ACTION project Magali.Barbier@onera.fr Outline Introduction Scenarios Scientific subjects Implementation Data fusion Decision: planning and supervision Evaluation Demonstration Conclusion & Perspectives 2/24 1
ACTION project - Introduction Funded by French Defence procurement agency DGA, 2007-2015 Consortium ONERA: The French Aerospace Lab LAAS/CNRS: Laboratory for Analysis and Architecture of Systems / French National Centre for Scientific Research Objectives: cooperation of multiple heterogeneous autonomous vehicles to upgrade the performance of the localization function Scientific experiments: air-ground and air-sea scenarios 3/24 ACTION project - Scenarios 4 air-ground scenarios, from 2 to 12 vehicles (AAV and AGV) Infrastructure surveillance mission: localization and tracking of non-cooperative targets Growing complexity: areas, targets, disruptive events 3 2 1 AAV Autonomous Aerial Vehicle AGV Autonomous Ground Vehicle 4/24 2
ACTION project - Scenarios 2 air-sea scenarios, with 2 and 3 vehicles (AAV, AUV, ASV) Securing mission Fight against water pollution mission AAV Autonomous Aerial Vehicle AUV Autonomous Underwater Vehicle ASV Autonomous Surface Vehicle 5/24 ACTION project - Scientific subjects Data fusion Decision making = planning + supervision software architecture distributed on each unmanned vehicle for the team to be autonomous Data: vehicle itself (proprio and exteroceptive sensors), other vehicles, mission operator Data fusion Mission Supervision Planning Actions: moves, perceptions, communications 6/24 3
ACTION project - Implementation 3-layers embedded implementation CDE: achievement of actions MONO: individual decisional layer MULTI: team decisional layer MULTI MONO CDE 7/24 ACTION project - Data fusion Data processing for the localization function SLAM (Simultaneous Localization and Mapping) of vehicles, targets and landmarks RT-SLAM: a generic and real-time visual SLAM implementation, C. Roussillon & al, 8th international conference on Computer vision systems, Sophia Antipolis (France), Sept. 2011 Layered environment modelling Tracking of targets 8/24 4
ACTION project - Decision Planning Coordination for cooperation: rendezvous, communication Off-line preparation and on-line repair or replanning Central role of supervision On-line control of the execution of planned actions Reactions trigger depending on current and predicted situations Execution of elementary actions: move, perception, communication, replanning process Interleaved functions Generates Domain + Problem Is used by Supervision Runs Planning Is used by Plan Generates 9/24 ACTION project - Decision Same hierarchical modelling of actions: Hierarchical Task Network HTN * Abstract and elementary tasks, pre- and post-conditions, several methods per tasks, sequenced or non-ordered methods Human expertise / flexibility Reparation process at the lowest level = involving the minimum number of vehicles, minimizing the communication * Erol, Hendler and Nau 1994 10/24 5
ACTION project - Planning Planning of vehicle actions = moves // perception + communication for rendezvous in a HTN format Requirements Communication not guaranteed regular rendezvous Temporal planning On-line repair and temporal flexibility Scenario I: Basic planner (N. Gobillot) Depth first tree search method in the non-instancied HTN Simple costs Selection of the first method Assignment of variables (vehicles, waypoints) Check of pre-conditions Check of post-conditions 11/24 ACTION project - Planning Scenario II: Constraint-based Planner * Use of InCell constraint-based local search library (Pralet &al, ICAPS 13) Takes into account time constraints and realistic transit times Plan = sequence of chunks Chunk = sequence of goto/observations for each vehicle External library for environmental model: reachability of zones, visibility for observation and communication, move durations Optimisation 1. Greedy procedure for building an initial plan 2. Local search to improve it (bestinsert, remove, permute, relocate) * Multi-Robot Planning and Execution for an Exploration Mission: a Case Study, G. Infantes & al, ICAPS Workshop on Planning and Robotics (PlanRob), 2014 12/24 6
ACTION project - Planning Scenario VI: Hierarchical Partial-Order Planner * POP = Set of actions + set of causal links + set of temporal constraints Algorithm: at each step, a plan is chosen, then a flaw, solving this flaw creates more plans to explore Principle: add abstract tasks to POP, that must be refined Improvements of the running time On-going work Reparation * HiPOP : Hierarchical Partial-Order Planning, P. Bechon & al, STAIRS-2014, symposium of ECAI, August 18-22, 2014, Prague (Czech Republic) 13/24 ACTION project - Supervision HTN computed by off-line planning: all vehicles + mission operator 14/24 7
ACTION project - Supervision HTN spread out on vehicles and for the mission operator, including communications/rendezvous 15/24 ACTION project - Supervision On-line reparation: Event = localized wreck 16/24 8
ACTION project - Evaluation Simulations Morse simulator based on Blender http://www.openrobots.org/wiki/morse/ Hybrid simulations with real vehicle/s in the loop Experimentations and demonstrations Esperce, military camp of Caylus Lake site from DGA 17/24 ACTION project - Demonstrations Vehicles ReSSAC aerial vehicles from ONERA Ground vehicles from LAAS Maritime vehicles from DGA 18/24 9
ACTION project - Demonstrations Air-ground scenarios I and III, Caylus Nominal mission with rendezvous Events management by the team: ground vehicle blocked, lost Video 19/24 ACTION project - Demonstrations Air-sea scenarios II et IV, DGA lake site IV = Fight against water pollution mission Phase 1 with aerial and surface vehicle to make a map of the pollutants (simulation!) in order to reduce the mission area Phase 2 with submarine vehicle to localize the wreck, the two other vehicles surveying the surface for eventual new pollutant localization 20/24 10
ACTION project - Demonstrations AAV itinerary ASV itinerary PHASE 1 Simulation RVs AAV itinerary ASV itinerary AUV itinerary PHASE 2 21/24 ACTION project - Conclusion Feasibility of a mission by a team of autonomous vehicles Added-value supplement existing means, decrease the workload of mission operators Autonomy: need to implement on-board the closed loop {perception situation assessment decision action} ACTION architecture distributed on each vehicle: Integrates the existing architecture (individual autonomy) Integrates the individual capabilities (moves, payload) Manages real communication constraints (neither continued, neither perfect) Supports localization (precise own moves, friends, target tracking) Generates environment models (off then on-line) Plans regular rendezvous (to maintain a common state of knowledge) Implements strategies to adapt to disturbing events, for communication recovery Supervises the planned actions and repairs when required 22/24 11
ACTION project - Perspectives The two demonstrations in 2015 (4 and 12 vehicles) Maturation of scientific work, new HIPOP planning algorithm New research avenues Cooperative target tracking Cooperative localization Replanning questions: where? who? Considers new disturbing events, studies new strategies Communications: which data to whom, when? Global situation assessment: estimation, prediction Team definition: hierarchy, real time update Role of the mission operator (authority sharing, conflict) 23/24 Thanks for your attention http://action.onera.fr/ 12