1. Physics for Games Programmers: Continuous Collision Detection Gino van den Bergen Playlogic Game Factory gino@dtecta.com

2. Overview  Discuss the GJK algorithm and show how to tailor it for detecting space-time collisions.  Discuss the Sweep and Prune algorithm and show how to tailor it for doing a broad phase in space-time.

3. Space-time Collision Detection  Perform collision detection in continuous space-time:  Construct a plausible trajectory for each moving object.  Detect collisions along these trajectories, using time as an added dimension.  Obtain expected time of impact and contact data for colliding pairs.

4. Plausible Trajectory?  Limited to trajectories with piecewise constant linear velocities.  Angular velocities are ignored. Rotations are considered instantaneous.

5. Only Translations  Solving continuous rotations is a lot more complex, so we’d rather dodge the issue.  Tunneling may occur for rotating objects, but is less visible and often acceptable.  Doing continuous translations only already fixes most of our problems and is doable in real time.

6. GJK Algorithm: Overview  An iterative method for computing the distance between convex objects.  First publication in 1988 by Gilbert, Johnson, and Keerthi.  Solves queries in configuration space.  Uses an implicit object representation.

7. GJK Algorithm: Pros  Extremely versatile:  Applicable to any combination of convex shape types.  Computes distances, common points, and separating axes.  Can be tailored for finding space-time collisions.  Allows a smooth trade-off between accuracy and speed.

8. GJK Algorithm: Pros (cont'd)  Performs well:  Exploits frame coherence.  Competitive with dedicated solutions for polytopes (Lin-Canny, V-Clip, SWIFT).  Despite its conceptual complexity, implementing GJK is not too difficult.  Small code size.

9. GJK Algorithm: Cons  Difficult to grasp:  Concepts from linear algebra and convex analysis (determinants, Minkowski addition), take some time to get comfortable with.  Maintaining a “geometric” mental image of the workings of the algorithm is challenging and not very helpful.

10. GJK Algorithm: Cons (cont'd)  Suffers from numerical issues:  Termination is governed by predicates that rely on tolerances.  Despite the use of tolerances, certain “hacks” are needed in order to guarantee termination in all cases.  Using 32-bit floating-point numbers is doable but tricky.

11. Configuration Space (recap)  A and B intersect: A – B contains origin.  Distance between A and B : length of shortest vector in A – B.

12. GJK Algorithm: Workings  Approximate the point of the CSO closest to the origin:  Generate a sequence of simplices inside the CSO, each simplex lying closer to the origin than its predecessor.  Terminate as soon as the current simplex is close enough.  In case of an intersection, the simplex contains the origin.

13. Support Mappings  A support mapping s A of an object A maps vectors to points of A, such that Any point on this face may be returned as support point

14. Primitives

15. More Primitives

16. Affine Transformation  Primitives can be translated, rotated, and scaled. For T(x) = Bx + c, we have

17. Convex Hull  Convex hulls of arbitrary convex shapes are readily available.

18. Minkowski Sum  Objects can be fattened by Minkowski addition.

19. Basic Steps (1/6)  Suppose we have a simplex inside the CSO…

20. Basic Steps (2/6)  …and the point v of the simplex closest to the origin.

21. Basic Steps (3/6)  Compute support point w = s A-B (-v).

22. Basic Steps (4/6)  Add support point w to the current simplex.

23. Basic Steps (5/6)  Compute the closest point v’ of the new simplex.

24. Basic Steps (6/6)  Discard all vertices that do not contribute to v’.

25. Affine Hull  The affine hull of points y i is the set of points x for which  The parameters λ i are better known as barycentric coordinates of x with respect to y i.  Point set Y is affinely independent if no y in Y is contained in the affine hull of Y \ {y}.

26. Closest Point of a Simplex  The point closest to the origin of a simplex:  Lies in the affine hull of the simplex.  Has non-negative barycentric coordinates.  Is an internal point of the simplex if all its barycentric coordinates are strictly positive.  Otherwise, the closest point is internal to the sub-simplex corresponding to the positive coordinates, or is a vertex.  As a vector, it is orthogonal to the affine hull of this sub-simplex (if it’s not a vertex).

27. Solving Barycentric Coordinates  The point x in the affine hull of points y i is closest to the origin if for an arbitrary y k  Thus, the vector x is orthogonal to n – 1 linearly independent vectors.  These vectors span the linear space corresponding to the affine hull.

28. Solving Barycentric Coordinates (cont'd)  After substituting and k = 1, we find the barycentric coordinates of point x by solving the following linear system:

29. Recursive Formula  Cramer’s rule and other linear-algebra wizardry yield a recursive formulation of the solution: Here, I X denotes the set of indices of points in X.

30. Recursive Formula (cont’d)

31. Johnson’s Algorithm  Compute the determinants for all 2 n -1 sub-simplices (including the full simplex).  The closest point of the simplex is the closest point of the unique sub-simplex X given by  All vertices not in X are discarded in the next iteration of GJK.

32. Termination (1/4)  As upper bound for the squared distance we take the squared length of v.  As lower bound we take v∙w, the signed distance of the supporting plane through w.  Terminate a soon as where ε rel is the tolerance for the relative error in the computed distance.

33. Termination (2/4) 0 

34. Termination (3/4)  In case of a (near) contact, v will approximate zero, in which case previous termination condition is not likely to be met.  As secondary termination condition we use where v is the current closest point, W is the set of vertices of the current simplex, and ε tol is a tolerance based on the machine epsilon.

35. Termination (4/4)  For objects that interpenetrate more extensively, GJK will eventually generate a tetrahedron containing the origin.  We may terminate as soon as Johnson’s algorithm returns a tetrahedron, since the tetrahedron must contain the origin.  In theory, the returned v should be zero, but normally it contains some rounding noise.

36. Numerical Issues (1/3)  Determinant computation suffers from cancellation when the simplex is needle-shaped.  Cancellation happens here:  Rounding errors may cause Johnson’s algorithm to fail: No proper simplex can be found.  The best thing to do is to simply terminate returning the last found distance.

37. Numerical Issues (2/3)  Noise in the computed v may result in failure to meet any of the termination conditions.  GJK cycles returning the same support point or oscillates between two support points.  Check whether the new support point is already a vertex of the current simplex (including discarded vertices).  Each support point should be added only once, so a returning support point is a sure sign of trouble.  Terminate as soon as this happens.

38. Numerical Issues (3/3)  In order to be absolutely positively sure that GJK terminates in all cases add the following check.  In theory, each new v should be shorter than the previous one, so terminate as soon as Here, ε flt is the machine epsilon of the floating- point format.

39. Separating Axes  If you only need to test for intersections then GJK may terminate as soon as v∙w becomes positive.  For positive v∙w, the vector v is a separating axis.  Separating axes can be cached and reused as initial v in future tests on the same object pairs.  When the degree of frame coherence is high, the cached v is likely to be a separating axis in the new frame as well.  An incremental version of GJK takes roughly one iteration per frame for smoothly moving objects.

40. Separating Axes (cont'd) 0   The supporting plane through w separates the origin from the CSO.

41. Shape Casting  Find the earliest time two translated objects come in contact.  Boils down to performing a ray cast in the objects’ configuration space.  For objects A and B being translated over respectively vectors s and t, we perform a ray cast along the vector r = t – s onto A – B.  The earliest time of contact is

42. Normals  A normal at the hit point of the ray is normal to the contact plane.

43. Ray Clipping

44. Ray Clipping (cont'd)  For, we know that  If v∙r > 0 then λ is a lower bound for the hit spot, and if also v·w > 0, the ray is clipped.  If v·r 0, then the ray misses.  If v·r = 0 and v·w > 0, the ray misses as well.

45. GJK Ray Cast  Do a standard GJK iteration, and use the support planes as clipping planes.  Each time the ray is clipped, the origin is “shifted” to λr,…  …and the current simplex is set to the last- found support point.  The vector -v that corresponds to the latest clipping plane is the normal at the hit point.

46. GJK Ray Cast (cont'd) The vector -v is the latest normal. The origin advances to the new lower bound.

47. Termination  The origin advances only if v·w > 0, which must happen within a finite number of iterations if the origin is not contained in the query object.  Terminate as soon as the origin is close enough to the query object, or we have evidence that the ray misses.  As termination condition in case of a hit we use

48. Accuracy vs. Performance  Accuracy can be traded for performance by tweaking the error tolerance ε tol.  A greater tolerance results in fewer iterations but less accurate hit points and normals.

49. Accuracy vs. Performance  ε tol = 10 -7, avg. time: 3.65 μs @ 2.6 GHz

50. Accuracy vs. Performance  ε tol = 10 -6, avg. time: 2.80 μs @ 2.6 GHz

51. Accuracy vs. Performance  ε tol = 10 -5, avg. time: 2.03 μs @ 2.6 GHz

52. Accuracy vs. Performance  ε tol = 10 -4, avg. time: 1.43 μs @ 2.6 GHz

53. Accuracy vs. Performance  ε tol = 10 -3, avg. time: 1.02 μs @ 2.6 GHz

54. Accuracy vs. Performance  ε tol = 10 -2, avg. time: 0.77 μs @ 2.6 GHz

55. Accuracy vs. Performance  ε tol = 10 -1, avg. time: 0.62 μs @ 2.6 GHz

56. Space-Time Sweep and Prune  Find pairs of objects whose axis-aligned bounding boxes overlap.  Take translations of objects over the time interval into account.  No false positives: box pairs must overlap at some point in the time interval.

57. AABB Computation  The projection of an object A onto an axis e is the interval:  Compute an object’s AABB by projecting the object onto the world axes.

58. Sweep and Prune (1/3)  For each world axis, maintain a sorted list of interval endpoints.  Maintain also the set of overlapping box pairs.  When a box moves, locally re-sort the lists by comparing and (if necessary) swapping the box endpoints with adjacent endpoints.

59. Sweep and Prune (2/3) Re-order endpoints of moving objects. B A

60. Sweep and Prune (2/3) Re-order endpoints of moving objects. B A

61. Sweep and Prune (3/3)  When swapping “][“ to “[]”, the intervals start to overlap.  When swapping “[]“ to “][”, the intervals cease to overlap.  If the intervals on the other axes overlap, then the box pair starts or ceases to overlap.

62. Implementation  “]” is greater than “[“ at equal positions  Mangle the least-significant bit: 0 = “[”, 1 = “]”.  Beware of float sign! For negative numbers: 1 = “[”, 0 = “]”.  Better still: use fixed-point numbers for storing endpoint positions.  Integer comparisons are generally faster than float comparisons.

63. Implementation (cont'd)  Endpoint lists are best implemented using arrays (std::vector):  Local sorting on arrays is CPU-cache friendly.  Each object maintains two indices per array.  Interval overlap tests can be done using array indices rather than endpoint positions.  Use a hash table or a dictionary (std::set) for storing the set of overlapping box pairs.

64. Adding Time: Encapsulation  Enlarge AABBs of moving objects, such that they encapsulate the swept AABBs.  Creates false positives: Encapsulating AABBs overlap where in space-time, the actual AABBs do not.

65. B Adding Time: Encapsulation (cont'd) Encapsulation results in false positives. A

66. Adding Time: Queued Swaps  Perform endpoint swaps in the proper order.  Calculate swap times and prioritize swaps on earliest time.  After each swap, re-evaluate swaps with new neighbors.  Needs a priority queue that offers a decrease-key operation (as in Dijkstra’s algorithm or A*).

67. Adding Time: Queued Swaps (cont'd) Swap endpoints in the proper order. B A

68. A Adding Time: Queued Swaps (cont'd) Swap endpoints in the proper order. B

69. Adding Time: Queued Swaps (cont'd) Swap endpoints in the proper order. B A

70. Performance  Queue operations dominate processing.  Despite added bookkeeping for decrease- key, best results so far are obtained using binary heaps. (Cormen et al.).  Binomial heaps turned out slower due to expensive delete-min operation.  Still have to try Fibonacci heaps, relaxed heaps, skew heaps, pairing heaps, …

71. Motion Coherence  Space-time Sweep and Prune is often faster than the original version when many objects are moving in the same direction.  Among a group of objects all having the same velocity vector not a single endpoint swap needs to be done.  Ideal for particle and fluid simulations.

72. Conclusion  Exact continuous collision detection of convex objects under translation is doable in real time.  Unfortunately, exact continuous collision response is impossible in real-time.  Gino’s paradox (in the spirit of Zeno): “Continuous motion is an illusion since you cannot resolve an infinite number of collisions in a finite amount of time.”

73. References  Thomas H. Cormen et al. Introduction to Algorithms, Second Edition. MIT Press, 2001.  E. G. Gilbert, D. W. Johnson, and S. S. Keerthi. A fast procedure for computing the distance between complex objects in three-dimensional space. IEEE Journal of Robotics and Automation, 4(2):192-203, 1988.  Gino van den Bergen. Collision Detection in Interactive 3D Environments. Morgan Kaufmann Publishers, 2004.

74. Thank You!  For papers and other information, check: www.dtecta.com  For information about Playlogic, please visit: www.playlogicgames.com