260 likes | 443 Views
Real-Time Multi-Agent Path Planning on Arbitrary Surfaces. Rafael P. Torchelsen 1 , Luiz F. Scheidegger 1 , Guilherme N. Oliveira 1 , Rui Bastos 2 , João L. D. Comba 1 1 Instituto de Informática - PPGC, UFRGS - Brazil 2 NVIDIA Corporation.
E N D
Real-Time Multi-Agent Path Planning on Arbitrary Surfaces Rafael P. Torchelsen1, Luiz F. Scheidegger1, Guilherme N. Oliveira1, Rui Bastos2, JoãoL. D. Comba1 1Instituto de Informática - PPGC, UFRGS - Brazil 2NVIDIA Corporation
Real-Time Multi-Agent Path Planning on Arbitrary Surfaces Obstacle Agent Goal Obstacle Shortest free path
Related Work Planar Surfaces 1 2 3 AvneeshSud, Erik Andersen, Sean Curtis, Ming Lin, and DineshManocha. Real-time Path Planning in Dynamic Virtual Environments Using Multi-agent Navigation Graphs. IEEE TVCG 2008 S. J. Guy, J. Chhugani, C. Kim, N. Satish, M. Lin, D. Manocha, and P. Dubey. ClearPath: Highly Parallel Collision Avoidance for Multi-Agent Simulation. ACM SIGGRAPH/Eurographics Symposium on Computer Animation 2009 AvneeshSud, Russell Gayle, Erik Andersen, Stephen Guy, Ming Lin, and DineshManocha Real-time Navigation of Independent Agents Using Adaptive Roadmaps. ACM Symposium on Virtual Reality Software and Technology 2007
Related Work Again: Planar Surfaces Distance computation is cheap Euclidian Distance: The mesh isn't considered on the distance computation Command and Conquer™
Contributions Multi-agent navigation on non-planar surfaces Multi-resolution distance computation GPU-based approach for collision avoidance CPU/GPU path planning pipeline
Outline • Related Work • Background • Global Navigation • Hierarchical Computation of Distances • Local Navigation • Collision Grid • Finding an Unobstructed Direction • Results • Discussion • Conclusion
Background Obstacle Agent Goal Obstacle Global Navigation Local Navigation
Global Navigation Several fields are need per frame Constant FPS Incremental Computation Fast Initial Movement Incremental Resolution Performance Parallel Computation Rafael P. Torchelsen, Francisco Pinto, Rui Bastos and João L. D. Comba. Approximateon-SurfaceDistanceComputationusingQuasi-DevelopableCharts. ComputerGraphicsForumv. 28 (PacificGraphics 2009)
Hierarchical Computation of Distances CPU GPU Collision Grid Path Planning Distance Field Computation Rendering Path Planning Distance Field Computation N Threads Path Planning M Threads Distance Field Computation ... Distance Fields Destination Resolution f4 r3 f1 r1 f4 r2 f3 r3 f2 r1 f2 r2 f3 r2 f3 r1 ... A1 A2 A3 A4 Am M Agents Distance Fields Destination Resolution f4 r3 f1 r1 f4 r2 f3 r3 f2 r1 f2 r2 f3 r2 f3 r1 To be Processed Distance Field Slot fd rr fd rr Not Ready fd rr fd rr ... Ready Available Last First
Outline • Related Work • Background • Global Navigation • Hierarchical Computation of Distances • Local Navigation • Collision Grid • Finding an Unobstructed Direction • Results • Discussion • Conclusion
Collision Grid • Advantage • 1 Agent <-> 1 Cell • Minimal Concurrent Access • Static Data Structure • Regular • GPU-Friendly Collision Grid A1 A2 A3 A4 Am M Agents ...
3-D Collision Grid: Advantage A1 A3 Obstacle A2
Collision Test An Agent ? ? ? pn Agentdestination p1 ? ? CP7 Path taken ? ? ? CP6 ? CPn Collisioncheckpointalongthe path CP5 ? ? ? ? ? ? ? Distancebetweencollisioncheckpoints r r CP4 ? ? ? ? ? CP3 A1 CP2 CollisiontestradiusaroundCP CP1 CP0 ? ? ? ? ? ? ? Cellchecked for collision
Collision Avoidance A3 p1 +2α +α +β A2 -α A1 -2α -β
Rotation Rotating the Gradient 0º 15º 45º 80º 90º 120º Goal A1 A1
Finding a Free Route A1 A1 A1 A2 A2 A2 A2 A1 p1 p1 p1 p1
Outline • Related Work • Background • Global Navigation • Hierarchical Computation of Distances • Local Navigation • Collision Grid • Finding an Unobstructed Direction • Results • Discussion • Conclusion
Results Tris: 20 000 Verts: 10 000 FPS: 30 Agents: 512 Grid Res: 1283 The agent color indicates the goal it is after After the agent reaches a goal it moves to another
Results Tris: 10 000 Verts: 5 000 FPS:30 Agents: 512 Grid Res: 1283 The agent color indicates the goal it is after After the agent reaches a goal it moves to another
Outline • Related Work • Background • Global Navigation • Hierarchical Computation of Distances • Local Navigation • Collision Grid • Finding an Unobstructed Direction • Results • Discussion • Conclusion
Discussion A1 A2 • Non Collision Free • First step towards the solution • Only collision avoidance • Collision Grid • Non smooth response • False positive
Conclusion • Multi-agent navigation on non-planar surfaces • Multi-resolution distance computation • GPU-based approach for collision avoidance • CPU/GPU path planning pipeline