Probabilistic Roadmaps

CS 326A: Motion Planning Probabilistic Roadmaps

The complexity of the robot’s free space is overwhelming

The cost of computing an exact representation of the configuration space of a multi-joint articulated object is often prohibitive • But very fast algorithms exist that can check if an articulated object at a given configuration collides with obstacles (more next lecture) •  Basic idea of Probabilistic Roadmaps (PRMs): Compute a very simplified representation of the free space by sampling configurations at random using some probability measure

Initial idea: Potential Field + Random Walk • Attract some points toward their goal • Repulse other points by obstacles • Use collision check to test collision • Escape local minima by performing random walks

But many pathological cases …

Illustration of a Bad Potential “Landscape” U q Global minimum

forbidden space Free/feasible space Probabilistic Roadmap (PRM) Space n

Probabilistic Roadmap (PRM) Configurations are sampled by picking coordinates at random

Probabilistic Roadmap (PRM) Sampled configurations are tested for collision

Probabilistic Roadmap (PRM) The collision-free configurations are retained as milestones

Probabilistic Roadmap (PRM) Each milestone is linked by straight paths to its nearest neighbors

Probabilistic Roadmap (PRM) The collision-free links are retained as local paths to form the PRM

g s Probabilistic Roadmap (PRM) The start and goal configurations are included as milestones

g s Probabilistic Roadmap (PRM) The PRM is searched for a path from s to g

Multi- vs. Single-Query PRMs • Multi-query roadmaps Pre-compute roadmap Re-use roadmap for answering queries • Single-query roadmaps Compute a roadmap from scratch for each new query

Sampling strategy This answer may occasionally be incorrect Procedure BasicPRM(s,g,N) • Initialize the roadmap R with two nodes, s and g • Repeat: • Sample a configuration q from C with probability measure p • If q  F then add q as a new node of R • For some nodes v in R such that v q do If path(q,v)  Fthen add (q,v) as a new edge of R 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 NoPath

Requirements of PRM Planning • Checking sampled configurations and connections between samples for collision can be done efficiently.  Hierarchical collision detection • A relatively small number of milestones and local paths are sufficient to capture the connectivity of the free space.  Non-uniform sampling strategies

PRM planners work well in practice Why?

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?

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 existsan 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. [Its goal is to minimize the expected number of remaining iterations to connect s and g, whenever they lie in the same component of F.] Why is PRM planning probabilistic?

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 this the case? So ...

Relation to Monte Carlo Integration f(x) A = a × b b (xi,yi) a x x1 x2

Relation to Monte Carlo Integration f(x) A = a × b b (xi,yi) a x But a PRM planner must construct a path The connectivity of F may depend on small regions Insufficient sampling of such regions may lead the planner to failure x1 x2

Visibility in F • 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}

F V(q) q Here, ε ≈ 0.18 ε-Goodness of F • 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 ε-good if every q in F is ε-good • Intuition: If F is ε-good, then with high probability a small set of configurations sampled at random will see most of F

F1 F2 Connectivity Issue

F1 F2 Lookout of F1 Connectivity Issue

F1 F2 Lookout of F1 Connectivity Issue The β-lookout of a subset F1 of F is the set of all configurations in F1 that see a β-fraction of F2 = F\ F1 β-lookout(F1) = {q  F1 | μ(V(q)F2)  βμ(F2)}

F1 F2 Lookout of F1 Connectivity Issue The β-lookout of a subset F1 of F is the set of all configurations in F1 that see a β-fraction of F2 = F\ F1 β-lookout(F1) = {q  F1 | μ(V(q)F2)  βμ(F2)} F is(ε,α,β)-expansiveif it is ε-good and each one of its subsets X has a β-lookout whose volume is at least αμ(X) Intuition: If F is favorably expansive, it should be relatively easy to capture its connectivity by a small network of sampled configurations

Comments • Expansiveness only depends on volumetric ratios • Itis not directly related to the dimensionality of the configuration space E.g., in 2-D the expansiveness of the free space can be made arbitrarily poor

Thanks to the wide passage at the bottom this space is favorably expansive Many narrow passages might be better than a single one This space’s expansiveness is worsethan if the passage was straight A convex set is maximally expansive,i.e., ε = α = β = 1

Experimental convergence g= Pr(Failure)  Theoretical Convergence of PRM Planning Theorem 1 Let F be (ε,α,β)-expansive, and s and g be two configurations in the same component of F. BasicPRM(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

Linking sequence

Theoretical Convergence of PRM Planning Theorem 1 Let F be (ε,α,β)-expansive, and s and g be two configurations in the same component of F. BasicPRM(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 For any ε > 0, any N > 0, and any g in (0,1], there exists αo and βo such that if F is not (ε,α,β)-expansive for α > α0 and β > β0, then there exists s and g in the same component of F such that BasicPRM(s,g,N) fails to return a path with probability greater than g.

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

Not really! Narrow passages are unstable features under small random perturbations of the robot/workspace geometry  Poorly expansive space are unlikely to occur by accident In retrospect, is this property surprising?

Most narrow passages in F are intentional … … but it is not easy to intentionally create complex narrow passages in F Alpha puzzle

PRM planners work well in practice. Why? • Why are they probabilistic? • What does their success tell us? • How important is the probabilistic sampling measure π?

good visibility poor visibility How important is the probabilistic sampling measure π? • Visibility is usually not uniformly favorable across F • Regions with poorer visibility should be sampled more densely(more connectivity information can be gained there) small lookout sets small visibility sets

g s Impact Gaussian [Boor, Overmars, van der Stappen, 1999] Connectivity expansion [Kavraki, 1994]

But how to identify poor visibility regions? • What is the source of information? • Robot and workspace geometry • How to exploit it? • Workspace-guided strategies • Filtering strategies • Adaptive strategies • Deformation strategies

Conclusion • 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

How important is the randomness of the sampling source? Sampler = Uniform source S + Measure π • Random • Pseudo-random • Deterministic[LaValle, Branicky, and Lindemann, 2004]

g s Choice of the Source S • Adversary argument in theoretical proof • Efficiency • Practical convenience

Probabilistic Roadmaps