On the Probabilistic Foundations of Probabilistic Roadmaps JeanClaude
On the Probabilistic Foundations of Probabilistic Roadmaps Jean-Claude Latombe Stanford University joint work with David Hsu and Hanna Kurniawati National University of Singapore D. Hsu, J. C. Latombe, H. Kurniawati. On the Probabilistic Foundations of Probabilistic Roadmap Planning. IJRR, 25(7): 627 -643, 2006. 1
Rationale of PRM Planners § The cost of computing an exact representation of a robot’s free space F is often prohibitive § Fast algorithms exist to check if a given configuration or path is collision-free § A PRM planner computes an extremely simplified representation of F in the form of a network of “local” paths connecting configurations sampled at random in F according to some probability measure 2
Procedure Basic. PRM(s, g, N) 1. 2. Initialize the roadmap R with two nodes, s and g Repeat: Sampling strategy a. Sample a configuration q from C with probability measure p b. If q F then add q as a new node of R Connection strategy c. For every node v in R such that v q do If path(q, v) F then add (q, v) as a new edge of R 3. 4. until s and g are in the same connected component of R or R contains N+2 nodes If s and g are in the same connected component of R then Return a path between them Else Return No. Path This answer may occasionally be incorrect 3
4
5
6
PRM planners work well in practice. Why? § Why are they probabilistic? § What does their success tell us? § How important is the probabilistic sampling measure p? § How important is the randomness of the sampling source? 7
Why is PRM planning probabilistic? § A PRM planner ignores the exact shape of F. So, it acts like a robot building a map of an unknown environment with limited sensors § At any moment, there exists an implicit distribution (H, h), where • H is the set of all consistent hypotheses over the shapes of F • For every x H, h(x) is the probability that x is correct § The probabilistic sampling measure p used by the planner reflects this uncertainty. The goal is to minimize the expected number of iterations to connect s and g, whenever they lie in the same component of F 8
Why is PRM planning probabilistic? § A PRM planner ignores the exact shape of F. So, it acts like a robot building a map of an unknown environment with limited sensors § At any moment, there exists an implicit distribution (H, s), where • H is the set of all consistent hypotheses over the shapes of F • For every x H, s(x) is the probability that x is correct § The probabilistic sampling measure p reflects this uncertainty. The goal is to minimize the expected number of remaining iterations to connect s and g, whenever they lie in the same component of F 9
So. . . § PRM planning trades the cost of computing F exactly against the cost of dealing with uncertainty § This choice is beneficial only if a small roadmap has high probability to represent F well enough to answer planning queries correctly [Note the analogy with PAC learning] § Under which conditions is the case? 10
Relation to Monte Carlo Integration f(x) But a PRM planner must construct a path The connectivity of F may on small A depend =a×b b regions (xi, yi) Insufficient sampling of such regions may lead the planner to failure x 1 a x 2 x 11
Visibility in F [Kavraki et al. , 1995] § Two configurations q and q’ see each other if path(q, q’) F § The visibility set of q is V(q) = {q’ | path(q, q’) F} 12
ε-Goodness of F [Kavraki et al. , 1995] § Let μ(X) stand for the volume of X F § Given ε (0, 1], q F is ε-good if it sees at least an ε-fraction of F, i. e. , if μ(V(q)) ε μ(F) § F is e-good if every q in F is e-good § Intuition: If F is ε-good, then with high probability a small set of configurations sampled at random will see most of F 13
Connectivity Issue F 1 F 2 14
Connectivity Issue F 1 The β-lookout of a subset X of F is the F 2 set of all configurations in X that see a β-fraction of FX β-lookout(X) = {q X | μ(V(q)X) β μ(FX)} Lookout of F 1 15
(ε, a, β)-Expansiveness of F The β-lookout of a subset X of F is the F 2 set of all configurations in X that see a β-fraction of FX F 1 β-lookout(X) = {q X | μ(V(q)X) β μ(FX)} Lookout of F 1 F is (ε, a, b)-expansive if it is ε-good and each one of its subsets X has a β-lookout whose volume is at least a μ(X) [Hsu et al. , 1997] 16
Comments § Expansiveness only depends on volumetric ratios § It is not directly related to the dimensionality of the configuration space ØIn 2 -D the expansiveness of the free space can be made arbitrarily poor ØIn n-D the passage can be made narrow in 1, 2, . . . , n-1 dimensions 17
18
Theoretical Convergence of PRM Planning Theorem 1 [Hsu, Latombe, Motwani, 1997] Let F be (ε, a, β)-expansive, and s and g be two configurations in the same component of F. Basic. PRM(s, g, N) with uniform sampling returns a path between s and g with probability converging to 1 at an exponential rate as N increases s g Linking sequence 19
Theoretical Convergence of PRM Planning Theorem 1 [Hsu, Latombe, Motwani, 1997] Let F be (ε, a, β)-expansive, and s and g be two configurations in the same component of F. Basic. PRM(s, g, N) with uniform sampling returns a path between s and g with probability converging to 1 at an exponential rate as N increases g = Pr(Failure) Experimental convergence 20
Theoretical Convergence of PRM Planning Theorem 1 [Hsu, Latombe, Motwani, 1997] Let F be (ε, a, b)-expansive, and s and g be two configurations in the same component of F. Basic. PRM(s, g, N) with uniform sampling returns a path between s and g with probability converging to 1 at an exponential rate as N increases Theorem 2 [Hsu, Latombe, Kurniawati, 2006] For any ε > 0, any N > 0, and any g in (0, 1], there exists ao and bo such that if F is not (ε, a, b)-expansive for a > a 0 and b > b 0, then there exists s and g in the same component of F such that Basic. PRM(s, g, N) fails to return a path with probability greater than g. 21
What does the empirical success of PRM planning tell us? § It tells us that F is often favorably expansive despite its overwhelming algebraic/geometric complexity § Revealing this property might well be the most important contribution of PRM planning 22
In retrospect, is this property surprising? § Not really! Narrow passages are unstable features under small random perturbations of the robot/workspace geometry 23
In retrospect, is this property surprising? § Not really! Narrow passages are unstable features under small random perturbations of the robot/workspace geometry § [Chaudhuri and Koltun, 2006] PRM planning with uniform sampling has polynomial smoothed running time in spaces of fixed dimensions (Recall that the worst-case running time of PRM planning is unbounded as a function of the input’s combinatorial complexity) § Poorly expansive space are unlikely to occur by accident 24
Most narrow passages in F are intentional … … but it is not easy to intentionally create complex narrow passages in F Alpha puzzle 25
PRM planners work well in practice. Why? § Why are they probabilistic? § What does their success tell us? § How important is the probabilistic sampling measure p? § How important is the randomness of the sampling source? 26
How important is the probabilistic sampling measure p? § Visibility is usually not uniformly favorable across F good visibility small lookout sets small visibility sets poor visibility § Regions with poorer visibility should be sampled more densely (more connectivity information can be gained there) 27
Impact s Gaussian [Boor, Overmars, van der Stappen, 1999] g Connectivity expansion [Kavraki, 1994] 28
§ But how to identify poor visibility regions? • What is the source of information? Ø Robot and workspace geometry • How to exploit it? Ø Workspace-guided strategies Ø Filtering strategies Ø Adaptive strategies Ø Deformation strategies 29
Workspace-Guided Strategies § Main idea: • Most narrow passages in configuration space derive from narrow passages in the workspace § Methods: • Detect narrow passages in the workspace • Sample more densely configurations that place selected robot points in workspace’s narrow passages Uniform sampling Workspace-guided sampling [Kurniawati and Hsu, 2004] 30
Filtering Strategies § Main Idea: • Sample several configurations in the same region • If a “pattern” is detected, then retain one of the configurations as a roadmap node (~ a more sophisticated probe) More sampling work, but better distribution of nodes Less time is wasted in connecting nodes § Methods: • Gaussian sampling • Bridge Test • Visibility PRM 32
Adaptive Strategies § Main idea: • Use previous sampling results to identify which regions to sample next Time-varying sampling measure p § Methods: • Connectivity expansion • Diffusion (EST, RRT, SRT) • Active learning 33
Deformation Strategies § Main idea: • Deform the free space to make it more expansive § Method: • Free space dilatation 34
Deformation Strategies § Main idea: • Deform the free space to make it more expansive § Method: • Free space dilatation [Saha et al, 2004] dilated free space Relation with Gaussian sampling 35
How important is the randomness of the sampling source? Sampler = Uniform source S + Measure p § Random § Pseudo-random § Deterministic [La. Valle, Branicky, and Lindemann, 2004] 36
Choice of the Source S § § Adversary argument in theoretical proof Efficiency (or lack of) Robustness (or lack of) Practical convenience (or lack of) 37
Efficiency s g corridor width 38
Robustness 39
Practical Convenience § Think about implementing the Gaussian strategy with deterministic sampling 40
Conclusion § In PRM, the word probabilistic matters. § The success of PRM planning depends mainly and critically on favorable visibility in F § The probability measure used for sampling F derives from the uncertainty on the shape of F § By exploiting the fact that visibility is not uniformly favorable across F, sampling measures have major impact on the efficiency of PRM planning § In contrast, the impact of the sampling source – random or deterministic – is small 41
How to improve PRM planning? § By improving the sampling strategy! § How? By answering : What is the right granularity of a PRM collision-checking probe? Ø Build roadmaps where nodes and edges are not fully tested and labeled by probabilities (e. g. , lazy collision checking strategies) Ø Adapt the granularity of the probe (e. g. , to get more information on the local shape of F) 42
Open Problems for PRM Planners § Free space made of multiple subspaces of different dimensionalities, like in legged locomotion on rough terrain or arm manipulation [Bretl and Hauser] 43
Open Problems for PRM Planners § Motion spaces of linkages with huge number of degrees of freedom Millipede-like robot with 13, 000 joints [Amato et al. , Apaydin et al. , Singhal et al. , Cortes et al. ] [Redon, 2004] 44
45
46
Connection Strategies § Which nodes to connect? § Which shapes of local paths to use? Ø Limit the number of connections: • Nearest-neighbor strategy • Connected component strategy Large impact Ø Increase expansiveness: • Multiple fixed shapes of local path [Amato 98] small impact • Local search strategy [Geraerts and Overmars, 2005, Isto 04] uneven impact 47
Comments § Poor expansiveness is caused by narrow passages § But narrow passages do not necessarily imply poor expansiveness 48
Comments § Many narrow passages can lead to a more expansive F than a single one 49
Comments § Windy passages are more difficult than straight ones 50
Comments § A convex free space is (1, 1, 1)-expansive 51
Experimental Results with Free Space Dilatation Time (s) [not including robot thinning] SBL (a) (b) (c) Alpha 1. 0 (d) (e) (f) SBL* (a) 12295 9. 4 (b) 5955 32 (c) 41 2. 1 (d) 863 492 (e) 631 65 (f) >100000 13588 52
- Slides: 52