Development of STRIPS for real-world robotics
The deliberative approach: pros and cons
Issues in planning research
Development of STRIPS for real-world robotics
Shakey
Developed at SRI; original design (1969):
- Used a theorem prover to construct plans. These were too
inefficient to find nontrivial plans.
- Designed to execute plans without monitoring their success or failure.
Almost all plans of any length failed at some point due to wheel slippage,
measurement errors, etc.
Next generation design (1971):
Layered system:
- High-level operators used for planning and problem solving.
- Detailed work of finding paths and moving objects moved from the general
problem-solving level down into special-purpose programs called
intermediate-level actions (ILAs).
- ILAs consist of complex routines of low-level actions (LLAs) for
controlling the physical robot.
For example, Shakey had an ILA called NavTo (i.e., "navigate to")
- could move the robot from one place to another within a room
- used A* algorithm to plan the path
- called LLAs to execute the path
- did limited path corrections along the way
In our kitchen robot example, our planning is done at a high-level. We
would need, for example, a "mix" ILA that would plan out the intermediate-
level details for mixing. These details would make use of LLAs that
would perform the actual robot control.
Shakey controlled by PLANEX:
- accepted goals from the user
- called STRIPS to generate plans
- executed plans by calling the appropriate ILAs.
- PLANEX kept track of current world state, comparing it to the preconditions
of each subsequence in the original plan.
- Would execute the shortest plan subsequence that led to a goal and whose
preconditions were satisfied. If no subsequence applicable, would call STRIPS to
make a new plan.
Errors explicitly modeled - as the robot moved, uncertainty about
its location increased. Upon reaching an uncertainty threshold, LLA would
call on the vision subsystem to provide a new position fix.
What has survived in modern designs:
- specialized components for low-level control and geometric reasoning
- centralized world model for planning
- compilation of (sub)plans to increase speed
- execution monitoring to handle unexpected problems
The deliberative approach: pros and cons
- can optimize performance relative to its model of the world
- requires relatively complete knowledge about the world
- requires strong assumptions about the world model: consistent, reliable,
certain
- plans may become invalid as the world changes
Issues in planning research
- operators with conditional effects make reasoning difficult
- order in which goals are tackled can have a significant effect on
efficiency
- most planners are ahistoric -- i.e., they don't learn from
experience