380 likes | 387 Views
Optimal Policies in Complex Large-scale UAS Traffic Management. David Sacharny & Tom Henderson. ICPS Taipei, Taiwan 7 May 2019. Futuristic Vision. (Slide from Jared Esselman ; UDOT). Commercial Use Cases. 3D Mapping, Video Collection Delivery (Amazon, etc.) Inspections
E N D
Optimal Policies in Complex Large-scale UASTraffic Management David Sacharny & Tom Henderson ICPS Taipei, Taiwan 7 May 2019 ICPS 2019
Futuristic Vision (Slide from Jared Esselman; UDOT) ICPS 2019
Commercial Use Cases • 3D Mapping, Video Collection • Delivery (Amazon, etc.) • Inspections • Data (Re)Transmission • Air Taxis • Investment 2017: $506M • 1000’s of flights per day ICPS 2019
Drone HW Investment ($B) ICPS 2019
Utah Urban Air Mobility Idea (Slide from Jared Esselman; UDOT) ICPS 2019
UDOT UAM (cont’d) (Slide from Jared Esselman; UDOT) ICPS 2019
UDOT UAM (cont’d) Proposal: Airwaysabove roadways. ICPS 2019
UAM: Need to Plan Flights ICPS 2019
Geospatial Intelligence:BRECCIA ICPS 2019
BRECCIA ICPS 2019
URBAN Implementation BRECCIA Agent • The BRECCIA Agent represents the core abstraction for all agents in the system. • Agents are distributed across specialized machines such as UAVs, mobile laptops, or high performance computers. • The inherited components of each BRECCIA agent enable an overall system that is dynamic and data-driven. BDI Engine :Mission Planner Uncertainty Reduct. Goal DB P.L. Logic Module Hadoop GeoWave Connector RRT* Planner Accumulo Specialized Functions GeoServer :Weather Monitor :UAV Manager :User Example Instantiations of the BRECCIA Agent ICPS 2019
Which Path To Take? What about Wind? What about Rain? ICPS 2019
Airway Corridors E.g., over Salt Lake City
Airspace Volumes Z Y -X X -Y -Z (b) Action Directions (a) Airspace Volumes
Learning Optimal Action Policy 4x4 Grid ICPS 2019
Bellman Equations where: U(s) is the utility of state s a is an action A(s) is the set of actions in state s R(s) is the reward for state s is a horizon coefficient ICPS 2019
State Representation state space: * 3 integer grid coordinates * 3 wind vector values (x,y,z) * 1 precipitation value * 1 temperature value ICPS 2019
State Representation:Reduced state space: * G = {1,2,3,4}: grid indexes * W = {0,1}: no wind; wind * P = {0,1}: no rain; rain * T = {0,1,2}: cold, normal, hot (temp) ICPS 2019
Actions A = {X, -X, Y, -Y, Z, -Z} * move in one of the coordinate directions ICPS 2019
Probabilistic State Transition ICPS 2019
Reward Function ICPS 2019
Value Iteration Algorithm From: Russell & Norvig ICPS 2019
Experiments • Start Location: 1,1,1 (index 1) • Goal Location: 4,4,4 (index 64) • Blocked Cell: 4,4,3 (index 60) • Can’t exit 4x4x4 • Preference for horizontal motion ICPS 2019
State Utilities and Path ICPS 2019
Convergence for Utilities Grid cell [3,4,4] [3,2,4] [2,1,1] ICPS 2019
Optimal Policies X: RIGHT Z: UP Y: BACK ICPS 2019
Optimal Policies ICPS 2019
Cell Travel Density ICPS 2019
Policies with Wind in Y No action in Y axis! No Wind Strong Wind in Y Direction ICPS 2019
Current Work: Get Data! ICPS 2019
Current Work: Testing!Deseret UAS ICPS 2019
Conclusions • Developed effective and efficient optimal policy method • Converted core BRECCIA system to work for UAS Traffic Management • allows communicating, autonomous agents • Cloud computing ICPS 2019
Current Development ICPS 2019
Large-scale Simulation • http://www.cs.utah.edu/~cem/uav/
Questions? ICPS 2019
UTAH UAV Fleet ICPS 2019
Acknowledgment This material is based upon work supported by the Air Force Office of Scientific Research under award number FA9550-17-1-0077 (DDDAS-based Geospatial Intelligence) ICPS 2019