20

Motion Planning for Deformable Robots: 

Motion Planning for Deformable Robots Serhat Tekin 11/7/2006

Motivation: 

Motivation Motion planning is a classical problem Mostly for rigid or articulated robots Deformable variants are recent Massive configuration space even for simple cases Existing methods not directly applicable

Motivation: 

Motivation Why need deformable robots? Applications in Industry CAD and virtual prototyping Computer generated animation Bioinformatics Computer-aided surgery

Outline: 

Outline Different approaches Physically-based Anshelevich et al. Rice University Geometry-based Bayazit et al. Texas A&M University Constraint-based Gayle et al. UNC at Chapel Hill Conclusion

Outline: 

Outline Different approaches Physically-based Anshelevich et al. Rice University Geometry-based Bayazit et al. Texas A&M University Constraint-based Gayle et al. UNC at Chapel Hill Conclusion

Physically-based Approach: 

Physically-based Approach Builds upon a similar framework introduced for elastic plates Lamiraux et al. Rice University An extension to PRM that takes deformation energy into account Volume deformations represented by a mass-spring lattice

Continuous Mechanical Model: 

Continuous Mechanical Model Uses the linear elastic physical model For a point v, energy density is defined by F is the matrix of partial derivatives of the deformation function γ evaluated at γ-1(v) The energy of γ is

Discrete Spring Model: 

Discrete Spring Model Approximates the continuous model Two types of springs between the masses Straight springs Angular springs Constant is picked according to type Discritized energy function

Volume Deformation: 

Volume Deformation Things to consider Grasp/Manipulation constraints on volume Restricts positions on some parts of the volume i.e. fix the positions of some point masses Energy minimization

Path Planning: 

Path Planning Uses PRM for path planning Local planner Interpolates between manipulation constraints to form a sequence of intermediate constraints Elasticity limits Constants to prevent unnatural deformations Plane strain limit: How much the material stretches locally Curvature limit: How much the material bends locally

Results (on an SGI R10000): 

Results (on an SGI R10000) Deformable cable with fixed end (32x3x3 lattice) 14.5 mins (average) Elastic pipe through a cube with an L-shaped hole (21x3x3 lattice) 8h 39mins

Outline: 

Outline Different approaches Physically-based Anshelevich et al. Rice University Geometry-based Bayazit et al. Texas A&M University Constraint-based Gayle et al. UNC at Chapel Hill Conclusion

Geometry-based Approach: 

Geometry-based Approach PRM extension, similar to the first method Deformations are not represented by physical means Aims at a reasonable-time limit with plausible deformations, rather than physical correctness

Overview: 

Overview Critical steps in the algorithm Roadmap construction Querying and deformation

Roadmap Construction: 

Roadmap Construction Need to estimate the edge weights Two different heuristics Shrinkable robots Use rigid robots with different scales Edge weight is the sum of “shrink factor”s of endpoints Allowing penetration Work in C-space to estimate penetration depth Sample n different C-space vectors (empirically, n = 20) If any sample is collision free, accept Minimum depth is used for the edge weight

Roadmap Construction: 

Roadmap Construction Shrinkable robot Penetration Swept volume of the path found

Query: 

Query Deformable robot used in the query phase Collisions must be avoided by deformations Configuration accepted if deformation energy below threshold Edges with higher weights are likely to fail, so test them first

Deformations: 

Deformations Bounding-box deformation Deformer pushes object into collision-free condition ChainMail deformation Similar to FFD Deforming boundary box vertex affects neighbors Free-form deformation (FFD) Only used for visualization Geometric deformation Deform the colliding portion directly Translate along surface normals

Deformations: 

Deformations Geometric deformation: a) Colliding configuration b) Intersecting polygons c) Deformed version Bounding-box deformation

Results: 

Results

Results: 

Results (for “narrow” scenario)

Summary: 

Summary Both methods are PRM extensions Differ in the way they handle roadmaps and deformations Physically-based Deformation taken into account during roadmap construction Deformations are physical simulations Geometry-based Robot treated as rigid during roadmap construction Deformations are geometric

Advantages/Disadvantages: 

Advantages/Disadvantages (+) Both methods offer a generalized framework to the problem Same deformation scheme can be used with a different randomized planner (-) Can handle only simple robots and environments First approach is computationally expensive Second one is not physically accurate

Outline: 

Outline Different approaches Physically-based Anshelevich et al. Rice University Geometry-based Bayazit et al. Texas A&M University Constraint-based Gayle et al. UNC at Chapel Hill Conclusion

Constraint-Based Motion Planning: 

Constraint-Based Motion Planning M. Garber and M. Lin, Constraint-based motion planning using Voronoi diagrams. Proc. Fifth International Workshop on Algorithmic Foundations of Robotics (WAFR), 2002 Reformulate the planning problem as a boundary value problem (BVP) Builds on similarity between BVP and Motion planning Map initial and goal configuration to boundary values Map motion into a constrained dynamics function Solvable through dynamical simulation

CBMP Goal: 

CBMP Goal To find a (near) minimal set of constraints which are sufficient to solve the problem Example: A 2D rigid robot in a simple environment The robot must: Reach a goal Avoid obstacles

Constraints: 

Constraints Hints at how the object should move Hard constraints: Must be satisfied at each step No penetration or intersection with obstacles Robot must stay within boundaries Articulated links must stay together Joint limits must be satisfied Soft constraints: Encourage a certain behavior Robot should follow a guiding path Robot should move towards the goal configuration Robot should avoid nearest obstacles

Overall Architecture: 

Overall Architecture

CBMP for Deformable Robots: 

CBMP for Deformable Robots DPlan: Builds upon CBMP Represent deformation as a list of constraints Represent energy minimization as a constraint Two stage approach Off-line roadmap generation Simple PRM for a point robot Possibly contains collisions Runtime path query by constrained dynamic simulation Performs deformation and local adjustments to path

Simulating Deformation: 

Simulating Deformation Represent deformation as a list of constraints Considerations Continuum representation Energy minimization Volume preservation Interaction with the environment

Continuum Representation: 

Continuum Representation Uses a simple Mass-Spring framework Computationally inexpensive Simple implementation and relatively easy interaction with the environment

Energy Minimization: 

Energy Minimization Robot energy function (defined by springs): k is spring constant, d is current distance, L is rest length Relax the case, i.e. allow small changes to the volume

Volume Preservation: 

Volume Preservation Relaxation Measures internal pressure variations Computes a pressure constraint force to adapt to changes in pressure Uses a simplified model based on the Ideal Gas Law Ideal Gas Law Force due to pressure on a surface

Adjusting Pressure: 

Adjusting Pressure Internal pressure constant defines the robot behavior The RHS constant of Ideal Gas Law (nRgT) is assigned by trial-and-error Low Pressure Medium Pressure High Pressure

Interaction: 

Interaction Hard constraints for interaction with environment Bounding-volume collision detection Assumes collision if robot is within a tolerance to an obstacle Applies impulses and repulsion forces at the affected masses Soft constraints for global behavior Path following

Deformation Step: 

Deformation Step Perform collision detection Handle collisions to enforce non-penetration constraints Accumulate spring forces Fs Compute the volume V of the object Set P = nRgT / V For each face f on the geometry Set Fp = PA For each vertex v of f Find the pressure forces on v by adding Fp divided by the number of faces incidental to v

Summary: 

Summary Builds upon CBMP Adds constraints for deformation, path following, and interaction with the environment Uses a simplified global path to help escape local minima while using CBMP to make local adjustments to ensure a collision-free path

Advantages: 

Advantages Allows for complex robots Computes physically-plausible deformations Performs sampling in low-degree of freedom space (i.e. workspace)

Limitations: 

Limitations Does not ensure a path will be found Cannot guarantee accurate deformations Restricted ability to represent robots with sharp edges Applicable only to closed robots Limited scalability

Results: 

Results Ball in cup Many spheres Cup - 500 Polygons Robot – 320 Polygons Spheres - 3200 Polygons Robot – 320 Polygons

Results: 

Results Walls with holes Walls - 216 Polygons per wall Robot – 720 Polygons

Results: 

Results Tunnel - 72 Polygons Robot – 720 Polygons Tunnel

Performance Results: 

Performance Results

Improving Performance: 

Improving Performance Support for complex environments FlexiPlan: Path Planning for Deformable Robots in Complex Environments (FlexiPlan) Builds upon DPlan by improving primary bottlenecks Guiding path improvements Simulation improvements

Improvements: 

Improvements More optimal global guiding path Samples along the medial axis of the workspace to create a path (Medial Axis PRM) Generalized Voronoi Diagram is another possibility Computed efficiently with GPUs Simulation improvements Mass-Spring simulation More stable (Semi-Implicit Verlet integration) Supports angular springs to counteract shearing Better collision detection scheme

Collision Detection: 

Collision Detection Dominating factor in running time DPlan only uses a bounding volume to remove unnecessary checks BVH is not a viable option Robot is often too close to obstacles in most scenarios BVH would not eliminate most tests and incur an update cost Speed up collision tests by 2.5D overlap test Set-based computation

2.5D Overlap Test: 

2.5D Overlap Test Based on CULLIDE Choose a viewing direction Check whether R is fully visible with respect to O along that direction Utilize GPU occlusion query

Reliable GPU Check: 

Reliable GPU Check Might miss overlaps due to pixel precision To prevent this Determine the size of a pixel Compute Minkowski sum of the obstacles and robot with a pixel Conservative, since may include pixels from geometry which does not overlap

Set-Based Computation: 

Set-Based Computation Maintains a PCS (Potentially Colliding Set) throughout computation Initially everything is in the PCS Uses overlap tests to remove obstacles from the PCS Do exact collision detection on the PCS If number of primitives is small, test all pariwise combinations Else, use bounding boxes for speed-up

CD Speedup: 

CD Speedup

Slide51: 

Catherization scenario Catheter: ~10K triangles Arteries: ~90K triangles

Results: 

Results

Video: 

Video

Summary: 

Summary Guiding Path Follows the medial axis of the workspace Spring-Mass Support for larger systems Catherization scenario has over 100,000 springs Greater stability Collision Detection GPU-based culling and set partitioning

Summary: 

Summary Advantages Scales better to complex scenes Introduces a planning specific CD algorithm Limitations Same planning restrictions as DPlan No definite path May not have accurate deformation Restricted to closed objects Setting constants for the simulation Requires a high-end graphics card

Conclusion: 

Conclusion The area itself is relatively new The first two approaches can be thought as pioneering work The last one takes novel approaches to planning and collision detection Current methods need to be extended for Complex robot shapes Articulated deformable robots Deformable obstacles Dynamic environments

References: 

References F. Lamiraux, L. Kavraki. Path planning for elastic objects under manipulation constraints. International Journal of Robotics Research, 20(3):188-208, 2001. E. Anshelevich, S. Owens, F. Lamiraux, L. Kavraki. Deformable volumes in path planning applications.IEEE Int. Conf. Robot. Autom. (ICRA), pp. 2290-2295, 2000. O. B. Bayazit, H. Lien, and N. Amato. Probabilistic roadmap motion planning for deformable objects. IEEE Int. Conf. Robot. Autom. (ICRA), 2002. R. Gayle, M. C. Lin, D. Manocha. Constraint-Based Motion Planning of Deformable Robots. International Conference of Robotics and Automation, 2005. R. Gayle, W. Segars, M. C. Lin, D. Manocha. Path Planning for Deformable Robots in Complex Environments. Robotics: Systems and Science, 2005.