1 toward real-time multi-model physical simulation laurent grisoni december, 9th, 2005 lifl (ustl),...

1

Toward real-time multi-model physical simulation

Laurent Grisoni

December, 9th, 2005

LIFL (USTL), INRIA Futurs/IRCICA, CNRS

2

Short CV

Dec. 1999: PhD thesis from University of Bordeaux I, Subject: multiresolution for geometric modeling (on splines, and implicits)

From Sept. 2000: assistant professor at Polytech’Lille engineering school

Sept 2003 Sept 2005: INRIA temporary full-time research position

Research advising:

- 2 completed PhD thesis (J. Lenoir and J. Dequidt)

- 2 on the run (A. Theetten and R. Vatavu)

- 6 Master thesis advised since 1997.


3

introduction

Main research context: « real-time » physical simulation Allow users to interact on virtual models that behave like real-life

ones

Real-time physical simulation for end-point applications involves several points:

- Collision detection and response- Mechanical modeling- Visual modeling (geometry, texturing, etc…)- Interaction with human (possibly haptic): not covered by the whole

of this expose (only interaction between mechanical systems here)

For designing a sophisticated model in a simulator, we usually see an object as a four-layered box: « multi-layered objects » [Meseure 02]

Each layer is a wide research subject in itself The question to know how each interacts with others raises many

interesting research problems (e.g. dynamic cutting?)


[Müller 04]

4

introduction

Medical simulation is an intrinsically complex application field:

big issue to provide complex protocols

simulation speed, visual and mechanical accuracy are crucial

Strong application potential (teaching, pre-op)


5

introduction

Some physical phenomena are known to be complex …


6

introduction

… but most are more complex than we usually perceive.


7

introduction

In physical simulation:

Each model has its own cost, and representation limits


8

introduction

For a given object, choice of model usually (implicitely) depends heavily on the context of use

As computer scientists, we would like:Flexibility,Modularity,Re-usability,Etc…,

not the case at all so far, and there is much more than coding here


9

introduction

How to overcome such limits for real-time physical simulation? important to understand each model potential, and limits provide as much flexibility to models as possible, through adaptivity propose mechanical systems that would be seen as black boxes, dynamically adapting their representation to context of use: multi-model objects


10

introduction

According to our perception of research activity in each part of the global problem:


Collision & response

Mechanical Visual

Model representation potential

Poor activity Active field Active field

adaptivity Active field Still emerging Active field

Multi-model Active field emerging Quite classical

(this array does not take into account non-separable aspects…e.g. cutting)

: presented contributions

11

overview

• Models in physical simulation

• Adaptivity

• Multi-model

• Conclusion and Future work


12

overview

Models in physical simulation

• Adaptivity

• Multi-model



13


Collision detection: 2 sub-classes of problems• Geometrical detection

– Huge litterature– Only simple primitives are accessible in real-time– Discrete, or continuous… time matters here.

• From detection to response– Not necessarily a force (e.g. exact contact handling [Cani 93])– Most often relies on interpenetration distance between objects – Difficult, so far, to overcome discrete nature of simulation


14

Mechanical modeling : 3 sub-class of problems

• From Forces to PDE: – Many different models: mass-spring,

particules, FEM, meshless…• Movement integration

– Several techniques available– Implies strong mathematical background

for contributing– Main trend: choose between stability,

accuracy, and computation cost• Constraints between objects

– Anywhere forces are no longer enough for simulation

– Main trend: choose between constraint respect and system stability



15


Visual modeling• Refers to geometry, texturation, real-time

rendering: once again, active research fields• From geometry point of view, usefullness of

models usually depend on object dimension:– 1D: splines, subdivision, implicits, generalized

cylinders– 2D: meshes, tensor-based splines, subdivision

surfaces, point-based reconstruction– 3D: implicits, volumic spatial partitionning,

particule-based density fields• Each has its pros and cons • More or less expensive• More or less topologically flexible (dynamic cut)

We think real-time physical simulator is a healthy context for research on geometry/texture manipulation, and real-time rendering in a more general way.


16

Contribution: fast visualisation of implicit using controlled blending and contact (Graphite 03, with F. Triquet)

• Blending graph represented as binary matrix, handled using dword bit-clusters

• For each primitive, operations on blending list is expressed as bitwise operations on dwords

• Adaptation of marching cube to contact control

Improvement of speed and visual result



17


Contribution: deformable 1D model (J. Lenoir PhD thesis)

• Deformable spline model (MS4CMS’02)– Based on Lagrange equations

– Provides real-time, continuous mechanics

• Set of constraints using Lagrange multipliers (Graphite 03)

– Sliding point constraints


18

overview


Adaptivity

• Multi-model



19

About terminology:• Blurry definition, throughout litterature, of words such as

« multiresolution », « adaptive », « hierarchical »• « level-of –detail » seems the most general one• « adaptive » can relate to basic methods that refine

representation where needed.• Two questions may help classify things:

– Is the studied problem taking advantage of the hierarchy to speed up the calculus? (yes: multiresolution; no: adaptive)

– does the representation have the same memory cost as the non-hierarchical version? (yes: multiresolution/hierarchical; no: level-of-details)

Another problem: no specific keyword distinction between time and space


adaptivity

20

Mechanical modeling • Many results on generating PDE on adaptive models

[debunne00, grinspun02]• Few on numerical integration for real-time:

– a lot on PDE solving using wavelets, based on Galerkin’s work on linear operators

– Not usable within real-time context…• To our knowledge, nothing on adaptive handling of

constraints ([Redon 05]?)


adaptivity

21

Collision detection

• Very classical to use bounding-box hierarchy (ex: k-DOP [Klosowski 98])

• Seen as the most practical way to overcome the cost of that stage

• Available on low-deformable volumes [James 04]

• Some results exist about anytime-interruptible collision detection schemes ([O’Sullivan 01])


adaptivity

22

Geometric modeling• Splines have important wavelet-

based framework for hierarchical manipulation

• Sweldens’ Lifting Scheme opens road to biorthogonal constructions for any interpolation scheme

• Subdivision curves and surfaces are intrinsically adaptive, and (theoretically) parameterized

Strong theoretical support, yet difficult so far to handle for real-time physical simulation skinning


adaptivity

23

adaptivity

Contribution: High performance Generalized Cylinders Visualization (Shape Modeling 03) (with D. Marchal)

• Work achieved during the ARC SCI (collaboration with EVASION project, and IRCAD)

• Tesselation achieved using vertex blending, combined with non-deformed surface primitive

• Provides significant improvement of surface drawing quality, and speed• First example of external control for guaranteed frame-rate


24

adaptivity

Contribution: adaptive 1D mechanical B-spline (Graphite 2005) (J. Lenoir PhD thesis)

• B-spline provide strong theoretical support for adaptivity (refinable basis functions, knot removal techniques)

• Mechanical model is adapted to B-spline refinement, and knot removal

Thread cutting provided at no-cost Interesting first example of a model that

semantically becomes fully independant from its resolution


25

adaptivity

Contribution: Time-critical deformation solids (CASA’05) (J. Dequidt PhD thesis, collaboration with D. Marchal)

• Adaptive octree physical simulation, used as FFD for animating complex meshes

• Control loop supervises the model for automatic tuning of model to computation cost (under some tolerance)

• From multi-layer objects point of view:

– adaptive collision, and mechanics– generic deformable model for

complex (topologically constant) shapes


26

overview

Models in physical simulation Adaptivity

Multi-model



27

Multi-model

Mechanical Modeling• Few yet about such use for

mechanical modeling• Some results about level-of-details

for animation [Carlson 97]

• Use of corotationnal [Felippa 00] for large deformations can be seen as a use of rigid rules into deformable

• Some work use rigid approximations when needed (e.g. explosions)

• [Müller 05] takes the best of each world between rigid and deformable


28

Multi-model

Collision detection

• Collision pipe-line [Zachman 01] is already multi-model

• No real study yet, to our knowledge, on dynamic switching from a collision algorithm to another one (e.g. discrete-continuous)


29

Multi-model

For visual modeling

• Level-of-detail is well-known for geometry handling: notion of cost function [Funkhouser 93]

• Some work on smooth transition between color texture, bump and displacement textures [Becker 93]


30

Multi-model

Contribution: Asynchronous interactive Physical Simulation (INRIA Tech. Report) (J. Dequidt PhD thesis)

• First software framework for such objects

• Each mechanical system lives in its own world

• Practical implementation of each system has no influence on the global software structure


31

overview

Models in physical simulation Adaptivity Multi-model

Conclusion and Future work


32

Conclusion and Future Work

Short-term directions of work:

- Surgical thread simulation

(A. Theetten thesis, co-advising with french CEA)

- Dynamic cutting

- First multi-model study

- What about GPU?


33


The SOFA Project:• Simulation Open-Framework Architecture• Co-initiated by the simulation group at CIMIT (MIT/Harvard

Hospital) and ALCOVE project• Currently joined effort of 4 research teams (CIMIT Sim. Group,

EVASION, EPIDAURE and ALCOVE)• Supported by the Odysseus european project• 3 INRIA engineers currently working on the project: P.-J.

Bensoussan, S. Fonteneau, F. Poyer

goal: provide modularity to real-time physical simulation Application designers should be able to exchange functionnal moduluses

(generic API for each bloc) Take advantage of future multi-processors architectures


34



The core is almost done.

We can now start to split work

into small activity groups, and

have other research teams

participate…

35



Mid and long-term research: we think real-time physical simulation would get much of: extend, as much as possible, model limits for representing real world

• Mechanical accuracy for real-time (what about computation time as a physical perception parameter??)

• Computation cost (the smaller, the better)• Topological control (e.g. cutting)• Get free from classical link between visual skinning and mechanical

model

provide flexibility to models: • Adaptivity anywhere possible• Dynamic control of such adaptivity (AI? Robotics?)

propose multi-model objects:• Evaluate each model limits • Propose ways to switch from a representation to another: how, and when

(AI?)• Provide software solutions for manipulation of such objects (collision,

constraints, etc…) (potentially much to get from software research community)

[F. Blondel, S. Karpf]

1 toward real-time multi-model physical simulation laurent grisoni december, 9th, 2005 lifl (ustl),...

Documents

model potential

choice of model

sophisticated model

introductionmedical

research position research

representation limits

wide research subject

interesting research