Adaptive%20Intelligent%20Mobile%20Robotics%20Leslie%20Pack%20Kaelbling,%20PI%20Artificial%20Intelligence%20Laboratory%20MIT

1
Adaptive Intelligent Mobile RoboticsLeslie Pack
Kaelbling, PIArtificial Intelligence
LaboratoryMIT

• Reinforcement Learning
• for Robot Control
• The goal is to make it easier for humans to
  program robots by using reinforcement learning
  and supplying a very high-level task description
  (in the form of a reward function) instead of
  explicit control instructions. Instead of
  learning from scratch, we show the robot
  interesting parts of the space by either
  driving it around with a joystick or providing it
  with an example solution. Neither of these need
  actually perform the task that we want to robot
  to learn. All that they have to do is to allow
  the learning system to observe interactions with
  the environment and the rewards that this
  interaction generates.
• Two Learning Phases
• Corridor-Following Task

• Optical-Flow Based
• Navigation
• We want to be able to control robots moving
  through dynamic worlds using information about
  how objects are moving in the world relative to
  us. We can get this information from optical
  flow.
• A Simple Navigation Strategy
• Sum the flow magnitudes in each half of the image

Phase One
Phase Two
A
R
O
A
R
O
Learning System
Learning System
?
?
Phase 1
Phase 2
Average training
Best possible