Finding Optimal Robot Motion


Bryce Willey

January 22th, 2018
The Rice University logo
Dee Faught was built a robotic arm from Rice Students The NASA Mars Exploration Rover (Artists Concept)

Benefits of Robots:

  • can assist people with disabilities/immobility

  • can complete tasks that are unsafe or impossible for humans



To reliably use robots in uncontrolled environments, teleoperation is the current best choice

Dee Faught has to control the robotic arm himself with a controller
  • Limits deployment to skilled operators

Automating these tasks allows widespread deployment

Motion planning is essential to automation.

Feasible Motion Planning

  • Finds a connected path from start to goal
  • Doesn't collide with any obstacles

Optimal Motion Planning

  • Assigns a cost to every path $$ C:\sigma \to \mathbb{R}$$
  • Finds a feasible path that minimizes the cost
  • Sampling Based Planners (Kavraki, 1994, LaValle, 1998)

  • Optimization Based Planners (Ratliff et al., 2009, Kalakrishnan et al., 2011)
Only finds feasible path 20% of the time with larger column

Currently, few rigorous comparisons in the literature

Planners:

  • Sampling: RRT-Connect, PRM, BIT*
  • Optimization: TrajOpt, CHOMP, GPMP2

TrajOpt = 6/8 scenes

CHOMP = 4/8 scenes

GPMP2 = 7/8 scenes

RRT-Connect = 8/8 scenes

PRM = 8/8 scenes, 87.5% of the tries

BIT* = 8/8 scenes

Planners:

  • Sampling: RRT-Connect, PRM, BIT*
  • Optimization: TrajOpt, CHOMP, GPMP2

Pros

  • Probabilistic Completeness
  • Very reliable
  • Can find optimal paths


Cons

  • Must smooth the path after finding it
  • Finding optimal paths is time consuming

Pros

  • 5-10x Faster
  • Quality, smooth paths


Cons

  • No guarantee of finding a feasible path
  • Not as reliable as sampling planners

Optimization planners do special 'tricks' to speed up planning

What really makes optimization planners faster?

  • Discretizes 3D space into a fine grid
  • Calculates distance and gradient from each cell to closest obstacle border
    • > 0 if the cell lies outside of the obstacle
    • < 0 if the cell is inside an obstacle
A signed distance field with it's 3D slice

Usually done precisely using collision libraries (FCL, Bullet, etc)

Can be approximated using signed distance fields

Approximating the robot using spheres makes getting collision information O(1)

A UR robotic arm
A UR robotic arm approximated with spheres

Abstract signed distance fields and compare speed ups in sampling planners


Test on a wider variety of planning problems


Use sampling and optimization together: quality motion and faster convergence

Possible sources of performance in optimization planners is under investigation

Best motion planner is dependent on your specifications