Transcript lis-5-99
Visual Robot Navigation
Low-level behaviors are built in or tuned with
reinforcement learning
Avoid obstacles using optical flow
Fixate and drive to a distant object
Map starts as a graph of views with behaviors on
the arcs
With further experience
Aggregate different views of the same place
Incorporate metric information
4/12/2016
1
Experimental Set-Up
Robot works in same virtual environment as
human subjects
No need for the head-mounted display!
Easy to simulate robot motion
Most development work in software-only
environment
Validation runs on real robot wearing head tracker
4/12/2016
2
Current Progress
Obstacle avoidance, wall-following using optical
flow from textured objects in virtual scene
Simple histogram-based view representation and
matching
Learn a route in the environment
Human drives robot through corridors with
joystick
Robot finds sequence of its behaviors that are
most consistent
Stores route as views connected by behaviors
Can re-create the route
4/12/2016
3
Technical Issues
Representation and matching of views
Pixel arrays vs. histograms
2D vs. 3D
Aggregation of views into places
Role of odometry
Role of 3D geometry
Role of metric information
Angles and distances on arcs vs. real 2D
embedding
4/12/2016
4