CSCE 221 Culture Assignment 3
« previous | Tuesday, May 3, 2011 | next »
Evaluation of Methods for Modeling Contact in Multibody Simulation
Dr. Dylan A. Shell
Rigid body Simulations: Applications of physics equations how do robots react to simulation environments?
We want our algorithms to be
- Fast
- Accurate (represent reality)
- Robust
Interactions between bodies
- Friction, other forces
Picking the best algorithm is tricky (challenging as it is important):
- A "best" algorithm is for a certain domain
Contact models: When and how are two bodies touching each other? What forces are at work to separate the objects.
Comparison of Algorithms
Primary focus on speed and robustness
What does it mean to be accurate?
- Very challenging issue historically
4 Different contact models (Impulse-based Event Driven)
- Newton Model
- Mirtich's Method
- Anitescu-Potra Method (3 solvers)
- Convex optimization method
7 benchmark scenarios:
- Impacting sphere colliding with ground plane (Netwon's was way off)
- Sliding Box
- Sticking Box
- Box Sliding with Circular Motion
- Stack of Boxes (Some models failed with 16 boxes (64 contact pointes)
- Mobile Robot Locomotion
- Manipulator Grasping (only Convex worked)
More resolution (accurate experiments) required more time
Conclusion
Simple problems:
- Anitescu-Potra is recommended
- Standard Lemke solver was fastest (and more robust than PATH)
Newton's contact model is very robust, but not very accurate
Mirtich's is fast, but doesn't work well with multiple contacts
Roadmap-Based Pursuit-Evasion in 3D structures
Jory Denny
Problem:
- Pursuers attempt to capture evaders
Applications:
- Search and rescue coverage
- Multi-level guarding
- Games, graphics, VR, robotics
Encoding environment:
- 2D or 3D
- Grid, Polygon, ROadmap
Agents:
- Communication
- Cooperation
- Behavior (rules that agent will follow) and roadmap
- View radius and angle
- Knowledge
Some examples get complicated very fast
Algorithm:
Get all agents, apply behavior
Pursuer:
Search for a target until found. Once found, develop strategy to capture probability of evader location can be stored in a graph
Evader:
Attempt to improve hiding location until evasive action necessary
Hiding location score: distance and direction to pursuers, visibility to pursuer (surrounding obstacles
Scenarios:
- Immobile
- Tracking
Basic vs. Frontier Search: Frontier searches faster
Probablistic vs. Basic Pursuing: Probablistic performs better (remembers location of evaders
Robust Recognition of Planar Mirrored Walls using a single view
Dr. Dezhen Song
Humans like shiny things. Mirror are shiny. Therefore humans like mirrors (modus ponens)
Robots can't detect mirrors.
Gallup's Mirror Test Chimpanzee's, Dolphins, and humans can. Gorillas and some birds cannot.
Curvature mirrors: planar mirrors are difficult to detect, but curved mirrors are easier
Polarization changes after reflection; easy to detect with stereo vision...
Given images containing stationary objects and relections, calculate the plane of the mirror.
Calculate mirror normal in 3D space: Normal is perpendicular to mirror plane, so calculate the cross product of the location vectors of the real and virtual object
Extracting real-virtual feature pairs with SIFT (128 dimensions