A Service Paradigm for Reconfigurable Agents

1
A Service Paradigm for Reconfigurable Agents

• Gary Holness, Deepak KaruppiahSubramanya Uppala,
  Roderic GrupenChandu Ravela

Laboratory for Perceptual RoboticsDepartment of
Computer ScienceUniversity of MassachusettsAmher
st, MA 01002USA
2
Topics

• Robotic Agent Physiology
• The general idea of The Service Model
• Sensor/Motor Service
• Tracking Example
• Future Directions

3
Robotic Agent physiology
sensors
environment
motors
objective

• Sensors/Motors interface to computational
  resource running an algorithm
  
• Progress guided by objective function
• Custom device protocols for sensors and motors
• Upgrading sensory motor suite means integrating
  new protocols

4
Service Paradigm

• Add assembler which knows agent structure
• Add computation at the site of sensors and
  motors
  
• Offer all pieces as tethered or non-tethered
  services across the network
  
• Services deliver code or can reach across the net
  as a proxy. The proxy hides protocol details.

5
Service Paradigm contd
Computational Resource
Assembler
Objective
Algorithm
Sensor
Motor

• Fault tolerance and adaptation through
  redundancy
  
• Resources necessary to accomplish a task can be
  swapped in/out based on quality metrics
  
• Can perform lazy loading of algorithms to meet
  finite computational budgets
  
• Our example swaps in/out sensory-motor resources

6
Sensor Service
Host/Embedded System
discovery/joinprotocol
filter
proxy
public interface Sensor extends Device
DeviceRegistration reportFeature(Type type,
long duration, long period,
RemoteEventListener target)
throws LeaseDeniedException,
RemoteException
7
Motor Service
Host/Embedded System
discovery/joinprotocol
proxy
public interface Actuator extends Device
DeviceRegistration actuate(ControlRef ref,
long duration, long period,
RemoteEventListener target)
throws LeaseDeniedException, RemoteException
8
Higher level services
This construction is recursive
A higher level service gets its raw information
from another service These services can be tether
ed or non-tethered At each level you can add comp
utation
9
Tracking Example

• Agent assembled by user (UI) at run time
• Redundant camera services for panoramic and
  pan-tilt-zoom. Higher level services extract
  headings from features.
  
• Motor service for asserting position references
  on pan-tilt-zoom camera
  
• Triangulation algorithm aggregates two heading
  sensors and a motor service
  
• Objective function rates resources based on
  liveness and reachability.
  
• Aquires and releases resources as it goes.

10
Future Directions

• Services for audio and mobile robots
• Downloadable mechanized assembly process which
  learns which assemblages are most successful and
  makes them persistent
  
• Load multiple versions of an algorithm
  (precision, energy, time optimal) and learn a
  policy for loading
  
• Switching among feature types and adjusting
  periods from the registration token
  
• Very long running (weeks vs. days) experiments
• Activity maps and learning the activity in a room