Kinematical Animation.

Transcription

1 Kinematical Animation

2 3D animation in CG Goal : capture visual attention Motion of characters Believable Expressive Realism? Controllability Limits of purely physical simulation : - little interactivity - high complexity for expressive characters

3 animation in 3D CG Two heritages Cartoons How to represent motion on a 2D visualization Robotics Mathematics foundation of 3D motion

4 Heritage from cartoon Sampled motion, film framework (24 images per second) Animation workflow from photographs (Muybridge, Marey)

5 Early Mathematics of motion Etienne-Jules Marey ( ) physiologist inventor of chronophotography (1882) cinematography invented in 1895 (L. Lumière) The graphical method : one motion can be represented by a curve

6 Heritage from robotics Chain of rigid articulations + Forward kinematics + Inverse-Kinematics algorithm + Motion planning Animation skeleton : appropriate degrees of freedom for expressivity

7 3D animation : interpolation+skeleton Keyframes and tangents

8 Interpolation Interpolation of a 1D-scalar value One function p(u) with unknown parameters, Typically cubic polynomial: p(u)=σ n=0..3 a n u n C (n) known constrains on points { u i, (d n p/du)(u i )= b i } 4 are enough for exact cubic polynomial Direct solving for a n p(u) is thus known for every u

9 Interpolation Practical case : Hermit polynomial (spline) C 1 continuity between sets of 2 points 2 positions and tangents at these positions p(u) = [ a 0 a 1 a 2 a 3 ] [ 1 u u 2 u 3 ] t = A t Q(u) p(0) = A t Q(0) = b 0 p (0) = A t Q (0) = b 1 p(1) = A t Q(1) = b 2 p (1) = A t Q (1) = b 3 A t [ Q(0) Q (0) Q(1) Q (1) ] = [ b 0 b 1 b 2 b 3 ] t A t Q 4x4 = B t => A t = B t Q -1 p(u) = B t Q -1 Q(u) p( u=(t-t0)/(t-t1) )

10 Interpolation Interpolating 3D rotation Canonic representation : SO(3) matrix but M 0,M 1 SO(3) > (1-α)M 0 +αm 1 SO(3) Represent SO(3) matrix with Euler angles any rotation in R 3 can be represented by 3 angles M = R x,ψ R y,θ R z,φ = Multiplication order matters! => Animation by interpolating angles

11 Interpolating 3D rotation Limitations of Euler angles Non-uniqueness of position and path [θ x,θ y,θ z ]=[0,0,0] [θ x,θ y,θ z ]=[0,0,0] [θ x,θ y,θ z ]=[π,0,0] M = [θ x,θ y,θ z ]=[0,π,π] M =

12 Interpolating 3D rotation Limitations of Euler angles Gimbal lock θx, [θ x,θ y,θ z ]=[0,0,0] [θ x,θ y,θ z ]=[0,π/4,0] [θ x,θ y,θ z ]=[0,π/2,0] 1 degree of freedom is lost : change in θ x change in θ z

13 3D rotation Axis-angle any rotation in R 3 is a planar rotation around an axis strong link with quaternion n v = R θ,n v = cosθ v + sinθ n v + (1-cosθ)(v n) n v θ v R θ,n = cosθ Id + sinθ [n] + (1- cosθ)nn t Rodrigues formula o

14 Quaternion H: Extension of standard complex q = [ s, x, y, z ] = s +ix + jy + kz with i 2 =j 2 =k 2 =ijk=-1 q* = [s, -x, -y, -z ] q 2 = qq* = s 2 + x 2 + y 2 + z 2 q -1 = q*/ q 2 q =1 q = [ cosθ, (sinθ)n ] H 1 with n R 3 and n =1 Any rotation in R 3 can be represented in H 1 x R 3, x = R θ,n x [0,x ] = q[0,x]q -1 with q = [ cos(θ/2), sin(θ/2)n ]

15 Interpolating 3D rotation Quaternion Linear interpolation in H does not work well q(t) = (1-t)*q 0 + t*q 1 Angular velocity is not constant q 0 q 1 Spherical linear interpolation is fine (SLERP) q 0 q 1

16 Interpolating 3D rotation Log and exp in H 1 q = [ cosθ, (sinθ)n ] = exp(θn) log(q) = [0,θn] q t = exp(t log(q) ) dq t /dt = logq q t dq t /dt = logq = [0,θn] = θ Application to SLERP SLERP(q0,q1,t) = q 0 (q 0-1 q 1 ) t SLERP(q0,q1,t) = (sin(ω - Ωt)q 0 + sin(ωt)q 1 )/sinω with cosω =q 0 q 1

17 Two fundamental cases Unconstrained motion waving arms, nodding head, etc => Forward Kinematics (FK) Constrained motion grasping an object, walking on the ground, etc => Inverse Kinematics (IK)

18 Forward Kinematics Direct application of 3D framework

19 Inverse Kinematics Articulated object Translational and rotational links Goal to reach M

20 Inverse Kinematics input : goal to reach (M) model parameter : Θ = ( θ 1, θ 2,, t 1, t 2, ), model parameters f(θ) position of kinematic chain end Find Θ /M=f(Θ ) 2 or 3 rotations: direct computation in R 2 or R 3 N articulations :?

21 Difficulties Two solutions : Range of solutions : No solutions :

22 f : direct kinematics Concatenation of matrix transforms f (Θ)= R 1 (θ 1 )T 1 (t 1 )R 2 (θ 2 ) T 2 (t 2 ) M 0 M 0 : position in rest pose (no rotations) Non linearity because of rotations

23 Zero of non-linear function Find Θ / f(θ)-m = 0 Linearization : given a current Θ and error E = f(θ)-m find h / E = f(θ+h) - f(θ) = f (Θ)h => h = f (Θ) -1 E => Θ := Θ + h iterate

24 Linearization Taylor series : Multivariate case : J Jacobian of f, linear from in h H Hessian of f, quadratic form in h

25 Jacobian Matrix of derivatives of several functions with severable variables : J:3xN matrix => not squared Use Pseudo-inverse for inversion : J + = J t (JJ t ) -1 if N>3 or J + = (J t J) -1 J t if N<3

26 Algorithm inversekinematics() { start with current Θ; E := target - computeendpoint(); for(k=0; k<k max && E > eps; k++){ J := computejacobian(); solve J h = E; Θ := Θ + h; E := target - computeendpoint(); } }

27 Joint limits Joint may have limits of variation For example realistic elbow is limited To enforce limits : test for limitation violation cancel parameter if violation in practice, remove column in J compute new J and J + compute new h

28 Adding constrains 1. if KerJ 0, degrees of freedom left to enforce a new constrains Ω JΩ = 0 2. if Θ solves JΘ = E, thus Θ+Ω is also solution J(Θ+Ω) = JΘ +JΩ = E + 0 = E 3. if general constrain C, need to project on KerJ: C p = (J + J I ) C check: J C p = J ( J + J I ) C = ( J J ) C = 0

29 Example: preferred angle Value : Θ pref Constraint C : C i = Θ i -Θ pref Modified algorithm : use h = J + E + (J + J-I)C => preserve convergence

30 Inverse Kinematics Other methods use J t instead of J + theory of infinitesimal works h = J t E use several 1D optimization Cyclic Coordinate Descent => faster but less accurate