Universal Plans for Reactive Robots in Unpredictable Environments

1
Universal Plans for Reactive Robots in
Unpredictable Environments

• By M.J. Schoppers
• Presented by Javier Martinez

2
Overview

• Integrate goal-directed advanced planning with
  sensor-driven reaction
• Allow the planner to generate new plans
  automatically

3
Motivation

• The linear approach used traditionally in AI has
  certain drawbacks such as
• Requires a lot of a-priori information
• Time consuming
• Delays actions arrival
• Additionally, the plans it produces cannot cope
  with unpredictable environments

4
Ideas

• Not committing to any particular sequence of
  events
• Let the environment dictate what to do next
• Planning is the goal-directed selection of
  reactions to possible situations
• If a situation satisfying condition P arises
  while trying to achieve goal G, then the
  appropriate response is action A

5
Planner Elements

• Primitive Actions are I/O conditions (in the
  context of a robot) that are maintained for an
  unspecified amount of time (i.e. speed up or slow
  down)
• Action Descriptions is the actual motions that
  comprise them
• Domain Constraints restrictions particular to a
  domain

6
Planner Elements

• The Goal is a condition to be achieved, instead
  of a world state

7
Universal plans

• Interpretation The plan has the shape of a tree
  and the interpreter traverses it by evaluating
  the environment at each node
• Hierarchy The idea is that a plan can become
  part of another as an action actions, thus being
  a sub-plan
• Competence Actions cannot replace planning even
  when both fulfill the same goal as plans are
  general while actions are conditioned

8
Plan Synthesis

• The process is done by back-chaining from the
  goal
• When back-chaining, what was a precondition now
  becomes a goal but in a negated form
• Back-chaining terminates when the preconditions
  are met or when a contradiction is found

9
Related Work

• Procedural Reasoning System (PRS)
• Reduced the amount of planning
• Behaviors are decomposed by hand
• Suffered from rigidity by not dealing with goal
  selection and rejection
• REX Project
• Continuously evaluates predicates
• Lacks symbolic representation
• Plans are hand-coded

10
Related Work

• Triangle Tables
• Create an index to a set of operators by
  extracting data from three sets
• Were the first ones to use the environment in the
  planning stage
• It suffers from the same rigidity as PRS

11
Advantages Limitations

• Being the plan basically a tree we can expect a
  computational efficiency of O(log(n))
• Approach limited to state spaces
• The approach is dependant on how fast sensors can
  deliver information to the plan interpreter

12
Paper Comments

• Goal directed planning along with reaction based
  behavior seem a more natural way of achieving
  goals
• Examples were difficult to follow and assumes too
  much knowledge about STRIPS operators

13

• Thank you!