Role Transfer

1
Role Transfer

• Human demonstrates a process (task, game)
• Annotates the demonstration vocally
• Robot initially participates vocally
• Robot uses vocalizations to segment percepts
• Goal robot incrementally takes over process
• Gracefully handles cooperative tasks for which
  human input will always be necessary

2
Mechanism

• Vocalizations treated as transitions of a finite
  state machine
• Perceptual state, filtered by attention system,
  recorded with transitions
• Supervised learning applied to find perceptual
  replacements for vocalizations
• Vocalizations can have histories that supply
  priors on learning
• Example big / small associated with
  appropriate visual filters in previous session
• Just need to be tweaked to the current situation
• Scheme can handle ambiguity, e.g. left / right

3
Protocol

• Vocalizations introduced before use
• Simplifies speech recognition
• Vocalizations used consistently
• The fewer states a vocalization is used in, the
  faster the process can be identified
• Robot can be prompted for prediction to test
  comprehension

4
Work in Progress
5
Goal Supervised learning
6
plus process context

• What is the difference between /y?lo/ and /l?ft/?
• Robot attends to same object in both
  circumstances
• Process provides additional context

7
Choice context
/y?lo/
/grin/
8
Response context
/l?ft/
/r?yt/
9
Future work
Improve percepts, action vocabulary Implement
contextual learning
Current system
Replace hard-coded behaviors Remove hard-coded
vocabulary
10
The Thane of Fife