SSS: A Hybrid Architecture Applied to Robot Navigation

1
SSS A Hybrid Architecture Applied to Robot
Navigation

• Jonathan H. Connell
• Presenter Lewis Girod

2
Introduction

• Describes a hybrid 3-layer architecture, SSS
• Describes a robot (TJ)
• 3-wheel base
• IR proximity sensors and sonar ranging
• TJs task is to
• build a map of office corridor environment
• accept commands to navigate using the map
• Describes implementation and presents an example
  of TJs performance

3
SSS is a hybrid architecture

• As we have discussed..
• Purely deliberative (SPA) architectures are slow
  because actuators run open-loop they cant
  respond to environmental dynamics
• Purely reactive architectures are fast and
  respond to environmental dynamics, but its hard
  to make them perform complex tasks
• Hybrid architecture
• reactive deliberative modules cooperate

4
SSS has three layers

• SSS has a three layer architecture
• Symbolic Goals are applied to this layer
• not real-time
• has complex representations data structures
• Subsumption Moment-to-moment operations
• real-time responses
• can produce non-linear discontinuous behavior
• Servo Tight control of actuators
• real-time, continuous and linear control systems
• smooth actuation

5
Questions to ask about TLAs

• How do the layers interact?
• How does each layer receive input from and
  manipulate the other layers?
• How messy is the middle layer?
• What is required to implement a different or more
  complex goal?
• Well try to address these along the way.

6
Three layers of SSS

• Three layers, three control techniques

Symbolic
event detectors
process parameterization
contingency tables
Subsumption
situation recognizers
setpoint selection
Servo
Sensors
Actuators
7
Servo layer

• Processes raw sensor data, drives actuators
• Continuous treatment of state and time
• Matched filters on raw sensor data
• recognizes situations, reports to subsumption
  layer
• Raw sensor data applied to tight control loops
• standard servo control system tries to maintain
  state variables at particular set points
• set points are configured by subsumption layer
• servo control eliminates jerky behavior by
  limiting acceleration

8
Subsumption layer

• Configures servo layer, reacts to situation
  input.
• Exports components of its state to symbolic layer
• treats time as continuous, state space as
  discrete
• Several behaviors run concurrently, with a rigid
  precedence hierarchy.
• Symbolic layer invokes behaviors from above
• Mechanics of behaviors and the significance of
  the state variables they export define this
  interface
• Contingency tables are bits of code
• implanted by symbolic layer to detect an event
  (analysis of internal behavior state) and to
  specify an immediate reaction

9
Subsumption layer

• Quick example of contingency
• TJs alignment behavior defines an interface.
• Based on odometry, a running average of TJs
  heading is maintained as a state variable in the
  subsumption layer
• An alignment behavior tries to steer TJ such that
  it drives in the direction of its average
  heading.
• Result symbolic layer can steer TJ by setting
  the heading variable and letting the alignment
  behavior drive
• Symbolic plans Down the hall 50 units, turn
  left
• Defines contingency when we have driven about 50
  units, look for a left turn if found, add 90 to
  avg. heading and signal event to symbolic level.

10
Symbolic layer

• Configures subsumption, responds to events
• enables and sets parameters for specific
  behaviors
• defines contingency tables, responds when they
  arise
• Contingency tables similar to sequencer
• treats time and state space as discrete
• Responds to high level input (goals) from users
• Maintains and uses complex representations
• TJ builds a coarse map of the environment
• uses it to plan paths from one point to another
• TJ can replan after each event.

11
Replan or preplan?

• Preplan Universal plans approach
• problem growth of contingency tree
• ex. Suppose a corridor is blocked.
• There are a large number of different places
  blockage could occur. Does it make sense to plan
  a contingency for each case?
• Replan Re-run the planner at each stage
• Coarse maps, navigation details in subsumption
  layer
• lower level of detail makes planning more
  tractable
• can do online local planning and extend plans
  offline
• Analogous to clever compression of universal plan

12
Details subsumption layer

• Subsumption supports a number of behaviors
• Obscacle avoidance (not described)
• Travel forward
• Slow down after n units have been travelled
• Wall following adjust heading
• Filtering to ignore minor bumps
• Alignment to average heading
• Wall following move closer to wall

High Priority Low Priority
13
Details symbolic layer

• Map representation
• coarse graph of landmarks and hallway segments
• landmarks are gaps in the wall (a room or
  corridor)
• map is constructed by matching repeated landmarks
• routes are planned using standard techniques
• To execute a plan,
• set up subsumption layer to move forward
• contingency table handles next turn flags event
• at the next event, replan if necessary

14
Results

• The results shown in the paper are good
• In their experiment they specified an initial
  route to compute the map, then requested a
  particular route
• It successfully navigated past various obstacles
• Not that much in the way of experiments
• just one experiment, run five times
• hard to conclude very much from this
• showing how to confuse the robot would be more
  useful

15
Analysis

• What about some harder tasks
• Blocked corridors (that are open on map)
• there are contingencies for being blocked
• but it would need to realize that the map had
  changed
• None of the tests exhibited hard cases at corners
• if the robot is not aligned at a corner it might
  overshoot
• this could happen if there was an obstruction
  there
• Doesnt handle localization (locating self on
  map)
• could use similar technique to matching in map
  construction
• turn on a wander behavior, match offline until
  located