Introducing a New Product - CR

Download Report

Transcript Introducing a New Product - CR

Hybrid architecture for autonomous
indoor navigation
Serge Belinski
Cyril Roussillon
Georgia Institute of Technology
CS 7630 – Autonomous Robotics
Spring 2008
Problem Statement
Autonomous navigation in a building
using an a priori map
and sonar sensors
Global planning:
A star
Algorithm
Graph best-first optimal path search
Heuristic = estimation of distance
A* optimal  heuristic admissible (lower bound)
e.g. euclidian distance
cost(S  G | A) ≥ dist(S  A) + heur(A  G)
Explores the most promising partial path
Algorithm
Initialization:
 Current node = start node
 Closed list = start node (nodes already considered)
 Open list = empty (nodes to consider, exploration front)
Nth step:
 Find neighbors of current node (no obstacles or closed list)
 For every neighbor:
If goal → end: path = parents
If in open list → update if better (cost and parent)
Else add in open list (cost and parent)
 Find the best candidate node in open list:
If open list empty → end: no solution
Else move from open list to closed list
set as current node
A* returns
Local obstacle avoidance:
Vector Field Histogram
Vector Field Histogram
Histogram Grid
Inspired by
certainty grids
increases one cell per
reading
accumulation of readings
creates certainty values
Vector Field Histogram
Polar Histogram
Restrained active
window
Angular obstacle
density
“Thresholded”
Vector Field Histogram
Adaptations
Maximum value for histogram
if robot stays still
Decrease histogram values
→ dynamic obstacles
A* and VFH
Global planning:
How to apply A*
A star
Modelization problems
Grid map → modelized as a graph
Usual way → immediate neighbors ...........
Problems:
Slow and memory-consuming for large grids
Gives low-level path
Want high-level path
Interpolation of discrete path does not
give optimal continuous path ....................
A star
Solution proposed
Neighbors = connectable by a straight line
without obstacle ………………….
Problems:
Graph of huge degree
Vicinity test pretty slow
Solutions:
Reduce the number of vertices
Precompute the graph
A star
Candidate intermediary points
Cells tangent to obstacles in convex parts
connect any pair of grid points
with a shortest path
A star
Characterization
Using a simple mask:
And the policy:
no purple cell obstacle
exactly one blue cell obstacle
at most one green “side” contains more than one
obstacle cell
A* and VFH
Testing
A*
Navigation points
= blue points
Dilation of
obstacle map for
embodiment
Demonstration
Small environment of two rooms simulated
With unknown static and dynamic obstacles
[Video]
Improvements
More and faster sonar → faster robot
Better localization than dead-reckoning for
large maps
Instability in the choice of the valley in VFH
Parameters tuning still improvable
Thank you!