Real-time Navigation of Independent Agents Using Adaptive Roadmaps Presenter: Robin van Olst
Feb 24, 2016
Real-time Navigation of Independent Agents
Using Adaptive Roadmaps
Presenter: Robin van Olst
The Authors
Avneesh Sud Russell Gayle Erik Andersen
Stephen Guy Ming Lin Dinesh Manocha
Elastic Bands – Quinlan and Khatib, 1993 Elastic Roadmaps – Yang and Brock, 2006
Real-time path planning for virtual agents in dynamics environments – Sud et al., 2006◦ Voronoi diagram generation using a GPU
Planning algorithms – LaValle, 2006◦ Random sampling
Self-organized pedestrian crowd dynamics and design solutions – Helbing, 2003◦ Local dynamics model (social forces)
Related work
Adaptive Elastic Roadmaps (AERO)◦ Global path planning method◦ Graph structure adapts to dynamic environments
Link bands◦ Local dynamics model◦ Augmented to AERO
Simulates a thousand of heterogeneous agents individually in real-time
Movie time!
Introduction
Adaptive Elastic Roadmaps (AERO)◦ Model description
Navigation with AERO◦ Link bands◦ Local dynamics model◦ Behaviour model
Implementation and results
Assessment
Outline
Based on a Generalized Voronoi diagram◦ Provides good initial clearance◦ Computes proximity information
Adaptive Elastic Roadmaps (AERO)
The Adaptive Elastic Roadmap◦ Consists of:
Milestones Links
Particles◦ Is a guiding path for agents
Find with graph search algorithms (A*)
Obstacles may block a path◦ Forces are applied to AERO
AERO Representation
Force on particles and milestones:
Internal forces:◦ Prevent unnecessary link deformation◦ Prevent roadmap drifting
External forces:◦ Respond to obstacle motion
AERO Force Computation
Necessary when a link is blocked Removal criteria
◦ Physics-based A link exceeds its stretching threshold
◦ Geometric-based The short distance to all obstacle is less than the
largest radius assign to an agent
AERO Link Removal
Repair removed links1. Check removed links2. Check disconnected milestones3. Repair is biased towards the area in the wake of
moving obstacles Lazy and incremental
Explore for new paths◦ Uses random sampling
Movie!
AERO Link Addition
Adaptive Elastic Roadmaps (AERO)◦ Model description
Navigation with AERO◦ Link bands◦ Local dynamics model◦ Behaviour model
Implementation and results
Assessment
Outline
Region of free space close to the nearest link◦ Space provides a collision-free path
Path planning◦ Starts at the nearest link◦ Each link is assigned a weight:
◦ Function of: link length, band width and the number of actors present on the band Each is weighted
High α: choose shortest paths (used for slow agents) High β: avoids narrow paths High γ: choose less crowded paths (used for aggressive agents)
AERO Link Bands
Local dynamics simulation◦ Helbing’s social forces model:
◦ Modified to add discomfort zones in front of moving obstacles Repulsive forces are biased along the motion of
obstacles
AERO Navigation: Link Bands
Agents can stand still, walk or jog◦ Depends on velocity◦ Uses non-parallel thresholds
Prevents oscillations◦ Aggressive agents prefer to jog
Higher maximum velocity
AERO Behaviour Modeling
Adaptive Elastic Roadmaps (AERO)◦ Model description
Navigation with AERO◦ Link bands◦ Local dynamics model◦ Behaviour model
Implementation and results
Assessment
Outline
3Ghz Pentium D CPU, 2GB RAM NVIDIA GeForce 7900 GPU, 512MB OpenGL
Optimizations◦ Spatial hash table of all entities and links
Efficient lookups and proximity computation◦ Voronoi diagram of all obstacles is computed
Scan a window to get all the obstacles within a certain range
AERO Implementation
Performance (in ms)
Cited in ‘Abnormal crowd behavior detection using social force model’ by Mehran et al.
Results
Adaptive Elastic Roadmaps (AERO)◦ Model description
Navigation with AERO◦ Link bands◦ Local dynamics model◦ Behaviour model
Implementation and results
Assessment
Outline
Adapts to dynamic obstacles◦ Handles changes in free space connectivity
Relates to real humans?
Able to simulate a thousand independently moving heterogeneous agents in real-time◦ Efficient
No assumptions on motion
Positive Points
Unrealistic high-DoF human motion◦ Only 3-DoF motion supported
Computed paths may not be optimal
Convergence is not guaranteed◦ Agents may get stuck in local minima
Limitations
Agents require a goal◦ No wandering
No grouping◦ Does not relate to real humans
Video shows rapid changes in orientation
Probably not able to simulate denser crowds
Negative Points
More efficient local dynamics model?
Complement method with:◦ Continuum Crowds’ discomfort fields◦ Navigational Fields’ directional preference
Suggestions