spatial neighbor learning in motion planning · from anc to anc-spatial each methods performs well...

1
Motion planning is concerned about finding a collision-free path from the start to the goal. A valid path is obtained from a sequence of valid configurations of the robot: a graph. Many methods connect samples in the free space in order to find the shortest valid path. An adaptive neighbor approach is a combination of connection methods that uses the appropriate method adaptively to every region in any environment. Spatial Neighbor Learning in Motion Planning PARASOL Lab: Department of Computer Science, Texas A&M University Diane Uwacu Chinwe Ekenna Nancy Amato [email protected] cekenna@ cse.tamu.edu [email protected] Introduction Conclusions Results and Analysis Spatial Learning We compare the ANC-Spatial with its predecessor ANC and several other individual methods. We major for the run-time to solve the query, the size of the roadmap and its connectivity. Our current goal is to prove that the ANC-Spatial can competitively adapt to heterogeneous environments. From ANC to ANC-Spatial Each methods performs well in a specific scenario. The ANC puts their strenghts together to solve heterogeneous problems. Why an Adaptive Approach Input. A connecting vertex q, a set of neighbor finders NF, a locality neighbor finder Nflocal, a local planner lp and a graph G Output. A connected graph G with additional edges. 1: Initialize Ninit = K set of Neighbors to q using Nlocal 2: Initialize a set of Neighbor finder probabilities Pq to the uniform distribution 3: for each n q Ninit do 4: Update Pq according to the record of < nf, reward, cost > stored in n 5: end for 6: Randomly select a nfq according to P N = nfq.Find NEIGHBORS(q,G) 7: for each n q N do 8 if lp.IS CONNECTABLE(q;n) then 9: G.ADD EDGE(q;n) 1 end if 11: end for 12: Update q and n with < nf, reward, cost > Input. A connecting vertex q, a set of neighbor finders NF, a local planner lp and a graph G Output. A connected graph G with additional edges Require: Let P be a set of probabilities such that pi is the probability of selecting nfi . Initialize pi = 1/|NF|, pi P1: Randomly pick nfi according to P 2: N = nfi.FIND NEIGHBORS(q,G) 3: for each n q N do 4: if lp.IS CONNECTABLE(q, n) then 5: G.ADD EDGE(q, n) 6: end if 7: end for 8: Let r be the success rate of lp over N 9: Let c be the cost incurred 10: Update (P, r, c) according to Equation 3 and Equation 4. ANC ANC_Spatial The ANC_Spatial updates connectors in every connected component. It updates the performance of connectors for every new connection made. To select the next connection method, we base our choice on the behavior in the local neighbors, instead of looking at the entire environment. For a node q that we want to add to the existing roadmap R, we find its n local neighbors in R. We find the cost and reward for the method used to connect n to R Based on the recorded information in n, we calculate new probabilities We then choose the best method based on the updated probabilities Protein Folding Movement Simulation in Building We have presented a new adaptive strategy to solve motion planning problems. It has shown promising results on both 2D and 3D environments, and handles heterogeneous environments with competence. We are currently investigating methods for applying our new methods to proteins and other environments with more degrees of freedom. Run time and Connectivity for the 2D maze The chosen connector is then used to add q to the existent roadmap R It is hard to know beforehand the best connection method to use This is more complicated when environments are heterogeneous Cost Number of coision decons made Reward Number of successl connecons over number of aempts made Existing roadmap R n local neighbors to q The probability is found based on the past performance of the connector (its weight). In ANC, we performed a global search each time we wanted to update the probability of the connector methods. In this spatial learning approach, we concentrate on localized learning. q References & Acknowledgments Adaptive Neighbor Connection for PRMs: A Natural Fit for Heterogeneous Environments and Parallelism. Chinwe Ekenna, Sam Ade Jacobs, Shawna Thomas, Nancy M. Amato. Hybrid PRM Sampling with a Cost-Sensitive Adaptivr Strategy. David Hsu, Gildardo Sanchez-Ante, Zheng Sun. Motion Planning for Robotic Manipulation of Deformable Linear Objects. Mitul Saha and Pekka Isto. This research was sponsored by Dr. Nancy Amato through Distributed Research Experiences for Undergraduates (DREU) I would like to thank my graduate student mentor Chinwe Ekenna, Shawna Thomas and the members of the PARASOL Lab for their support through this research. ANC vs _Spatial Which one gets is better? Method Scaled euclidean K-Rand R_Closest LPSwept ANC 0.054 0.298 0.648 ANC_Sp atial 0.163 0.837 0.0006 Final probabilities assigned to the three connection methods used in our experiments are shown below. ANC ranking: 1. BFNF2 2. Rr_kc 3. BFNF _Spatial rank: 1. Rr_kc 2. BFNF 3. BFNF2 ANC_Spatial assigns probabilities that reflect the real performance of the methods.

Upload: others

Post on 18-Nov-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Spatial Neighbor Learning in Motion Planning · From ANC to ANC-Spatial Each methods performs well in a specific scenario. The ANC puts their strenghts together to solve heterogeneous

●  Motion planning is concerned about finding a collision-free path from the start to the goal. ●  A valid path is obtained from a sequence of valid configurations of the robot: a graph.

●  Many methods connect samples in the free space in order to find the shortest valid path. ●  An adaptive neighbor approach is a combination of connection methods that uses the

appropriate method adaptively to every region in any environment.

Spatial Neighbor Learning in Motion Planning PARASOL Lab: Department of Computer Science, Texas A&M University

Diane Uwacu Chinwe Ekenna Nancy Amato [email protected] [email protected] [email protected]

Introduction

Conclusions

Results and Analysis

Spatial Learning

We compare the ANC-Spatial with its predecessor ANC and several other individual methods. We major for the run-time to solve the query, the size of the roadmap and its connectivity. Our current goal is to prove that the ANC-Spatial can competitively adapt to heterogeneous environments.

From ANC to ANC-Spatial

Each methods performs well in a specific scenario. The ANC puts their strenghts together to solve heterogeneous problems.

Why an Adaptive Approach

Input. A connecting vertex q, a set of neighbor finders NF, a locality neighbor finder Nflocal, a local planner lp and a graph G Output. A connected graph G with additional edges. 1: Initialize Ninit = K set of Neighbors to q using Nlocal 2: Initialize a set of Neighbor finder probabilities Pq to the uniform distribution 3: for each n ≠ q ∈ Ninit do 4: Update Pq according to the record of

< nf, reward, cost > stored in n 5: end for 6: Randomly select a nfq according to P

N = nfq.Find NEIGHBORS(q,G) 7: for each n ≠ q ∈ N do 8 if lp.IS CONNECTABLE(q;n) then

9: G.ADD EDGE(q;n) 1 end if 11: end for 12: Update q and n with < nf, reward, cost >

Input. A connecting vertex q, a set of neighbor finders NF, a local planner lp and a graph G Output. A connected graph G with additional edges Require: Let P be a set of probabilities such that pi is the probability of selecting nfi . Initialize pi = 1/|NF|, ∀pi ∈ P1: Randomly pick nfi according to P 2: N = nfi.FIND NEIGHBORS(q,G) 3: for each n ≠ q ∈ N do 4: if lp.IS CONNECTABLE(q, n) then 5: G.ADD EDGE(q, n) 6: end if 7: end for 8: Let r be the success rate of lp over N 9: Let c be the cost incurred 10: Update (P, r, c) according to Equation 3 and Equation 4.

ANC

ANC_Spatial

The ANC_Spatial updates connectors in every connected component. It updates the performance of connectors for every new connection made. To select the next connection method, we base our choice on the behavior in the local neighbors, instead of looking at the entire environment.

For a node q that we want to add to the existing roadmap R, we find its n local neighbors in R. We find the cost and reward for the method used to connect n to R

Based on the recorded information in n, we calculate new probabilities We then choose the best method based on the updated probabilities

Protein Folding Movement Simulation in Building

•  We have presented a new adaptive strategy to solve motion planning problems. •  It has shown promising results on both 2D and 3D environments, and handles heterogeneous environments with competence. •  We are currently investigating methods for applying our new methods to proteins and other environments with more degrees of

freedom.

Run time and Connectivity for the 2D maze

The chosen connector is then used to add q to the existent roadmap R

•  It is hard to know beforehand the best connection method to use •  This is more complicated when environments are heterogeneous

Cost Number of collision detections made Reward Number of successful connections over number of attempts made

Existing roadmap R

n local neighbors to q

The probability is found based on the past performance of the connector (its weight).

In ANC, we performed a global search each time we wanted to update the probability of the connector methods. In this spatial learning approach, we concentrate on localized learning.

q

References & Acknowledgments •  Adaptive Neighbor Connection for PRMs: A Natural Fit for Heterogeneous Environments and Parallelism. Chinwe Ekenna, Sam Ade Jacobs, Shawna Thomas, Nancy M. Amato. •  Hybrid PRM Sampling with a Cost-Sensitive Adaptivr Strategy. David Hsu, Gildardo Sanchez-Ante, Zheng Sun. •  Motion Planning for Robotic Manipulation of Deformable Linear Objects. Mitul Saha and Pekka Isto. •  This research was sponsored by Dr. Nancy Amato through Distributed Research Experiences for Undergraduates (DREU) •  I would like to thank my graduate student mentor Chinwe Ekenna, Shawna Thomas and the members of the PARASOL Lab for their support through this research.

ANC vs _Spatial Which one gets is better?

Method Scaled euclidean

K-Rand R_Closest

LPSwept

ANC 0.054 0.298 0.648 ANC_Spatial

0.163 0.837 0.0006

Final probabilities assigned to the three connection methods used in our experiments are shown below.

ANC ranking: 1.  BFNF2 2.  Rr_kc 3.  BFNF _Spatial rank: 1.  Rr_kc 2.  BFNF 3.  BFNF2

ANC_Spatial assigns probabilities that reflect the real performance of the methods.