The Stochastic Motion Roadmap: A Sampling Framework for Planning with Markov Motion Uncertainty Changsi An You-Wei Cheah
Introduction
Problem Sources of uncertainty Imprecision of execution of action a narrow passageway is unlikely to be robust to motion uncertainty
Failure other than collision Nonholonomic constraint Irreversible deflection in the way a deflection in the path due to motion uncertainty can result in failure to reach the goal
Countermeasure in the paper Stochastic Motion Roadmap Probabilistic method Explicitly consider the motion uncertainty Increase success rate of achieving the goal Characteristic Decide the most promising next move at each step Rely on percepts(feedbacks) to know current state
Build Road Map Probabilistic Road Map (PRM) Stochastic Motion Roadmap (SMR) s g SMR: Enhanced PRM Describe the possibility of transition from one state to another PRM Capture the connectivity
Stochastic Motion Roadmap s g Motion uncertainty
Build SMR Same as PRM isCollisionFree(x) isCollisionFreePath(x, y) Sample configurations Motion emulator with uncertainty getTransitions(x, u) Environment Describer x, y : configurations x, y u: control
GetTransitions(x, u)
Sampled next states: from continuous configuration space x Voronoi Graph
GetTransitions(x, u) Sampled next states x V1V1 V2V2 V3V3 V 1, V 2, V 3 ∈ C InControlNextProb xμV1V xμV2V xμV3V
GetTransitions(x, u) Sampled next states x V1V1 V2V2 V3V3 V 1, V 2, V 3 ∈ C Obstacles InControlNextProb xμV1V xμV2V xμV3V
V 1, V 2, V 3 ∈ C Obstacles InControlNextProb xμV1V xμV2V xμV3V
V 1, V 2, V 3 ∈ C Obstacles InControlNextProb xμV1V xμV2V xμOBTC0.15..
V 1, V 2, V 3 ∈ C Obstacles InControlNextProb xμV2V xμOBTC V1V1 V2V2 V3V3
V i, X i ∈ C μ i ∈ Μ Transition Probability Matrix InControlNextProb x1x1 μ1μ1 V2V x1x1 μ1μ1 OBTC0.33 x1x1 μ2μ2 VaVa PaPa x1x1 μ2μ2 VbVb PaPa.. x2x2 μ1μ1 VcVc PcPc The table means: when currently at state In, if control Control is taken, there is Prob chance to move to state Next
Query Goal At configuration i ∈ C, choose μ i ∈ Μ maximize the success probability
Bellman equation State Evaluation Compute recursively with BE, If j is within destination area, If j will lead to collision with obstacle, If otherwise Employ the idea of Dynamic Programming to memorize the intermediate results of J*(j)
Circle in Recursion Cyclic transition probability graph Utility function: penalty penalty
Stochastic Road Map Strengths Maximize the success rate for nonholonomic systems High fault tolerance from dynamic decisions General framework, very flexible for modeling the uncertainty Drawbacks Rely on an accurate percepts of current state Omit goal and obstacle dynamics Optimality restricted by magnitude of discrete representatives of CSpace
SMR for Medical Needle Steering Steerable needles are controlled by 2 degrees of freedom: Insertion distance Bevel direction Workspace is extracted from a medical image Obstacles are tissues that should not be cut by the needle
Bang bang steering car model State of the car is represented by a 4 dimension state space, s i = (x i, y i, θ i, b i ) Bevel direction of needle can be set to point left (b = 0) or right (b = 1)
SMR Implementation of bang bang steering car model A car moves δ between sensor measurements of states The set U consists of two actions: move forward turning left (u = 0), or move forward turning right (u = 1). As the car moves forward, it traces an arc of length δ with radius of curvature r and direction based on u. r and δ are random variables from δ ~ N (δ 0, σ δa ) & r ~ N (r 0, σ ra ), a ∈ {0, 1}
SMR Implementation of workspace Workspace: Rectangle of width x max and height y max. Obstacles: Polygons in the plane Zero-winding rule is used to detect obstacles. distance(s 1, s 2 ) = √[(x 1 − x 2 ) 2 +(y 1 − y 2 ) 2 +α(θ 1 −θ 2 ) 2 ] + M, where M → ∞ if b 1 ≠ b 2, and M = 0 otherwise. CGAL implementation of kd-trees is used to calculate fast nearest-neighbor Goal T ∗ as all configuration states within a ball of radius t r centered at a point t ∗.
Evaluation of SMR p s improves as the sampling density of the configuration space and the motion uncertainty distribution increase As n and m increase, p s (s) is more accurately approximated over the configuration space, resulting in better action decisions. Difficult problem: p s effectively converges for n ≥ 100,000 and m ≥ 20
Evaluation of SMR
T HANK YOU