Learning Prepositions for Spatial Relationships in BOLT Soar Workshop 2012 James Kirk, John Laird 6/21/2012 1
Dec 27, 2015
1
Learning Prepositions for Spatial Relationships in BOLT
Soar Workshop 2012James Kirk, John Laird
6/21/2012
2
Outline
• Prepositions in BOLT• Compositional Spatial Relationships• Learning Process in Architecture with SVS• Projection• Conclusions
3
Prepositions/Spatial Relationships in BOLT
• “What is to the right of the pantry?”• “Put the triangle in the stove.”• “Pick up the red object to the right of the pantry.”
4
Spatial Relationships as Compositions
y-alignedz-alignedx-greater than
Z
Y
X
“right of”
“Intersecting”y-alignedz-alignedx-aligned
“Above”z-greater than
Other potential compositions:
“In front of”y-less thanz-alignedx-aligned
5
Learn new prepositions that map to spatial relations “The red triangle is right of the blue sphere”
6Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Prep Learning
Action Knowledge
Procedural Memory
Preposition – Spatial RelMapping
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 1Perception
- Receive object information from
input link
- Create internal representations
Working Memory
Obj1- bl3- sp7
Obj2-re23-tr17
Obj3- bl3- sq8
Human: “The red triangle is right-of the blue sphere” Word – Category
Mapping
7Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Prep Learning
Action Knowledge
Procedural Memory
Preposition – Spatial RelMapping
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 2Message Parsing
- Parse the given sentence
- Create usable message
structures
“The red triangle is right-of the blue sphere”
Working Memory
LGSoar
Obj1- bl3- sp7
Obj2-re23- tr17
Obj3- bl3- sq8
relation-message ^word right-of ^o1 red, triangle ^o2 blue, sphere
Human: “The red triangle is right-of the blue sphere” Word – Category
Mapping
8Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Prep Learning
Action Knowledge
Procedural Memory
Preposition – Spatial RelMapping
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 3Interaction
Stack
- Designate message purpose
as preposition learning
Working Memory
LGSoar
relation-message ^word right-of ^o1 red, triangle ^o2 blue, sphere
Interaction
Prep. LearningObj1- bl3- sp7
Obj2-re23tr17
Obj3- bl3- sq8
Human: “The red triangle is right-of the blue sphere” Word – Category
Mapping
9Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Prep Learning
Action Knowledge
Procedural Memory
Preposition – Spatial RelMapping
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 4Preposition
Learning
Working Memory
Interaction
relation-message ^word right-of ^o1 red, triangle ^o2 blue, sphere
Obj1- bl3- sp7
Obj2-re23-tr17
Obj3- bl3- sq8
Prep. Learning
Prep Learning
right-ofred, triangle
blue, sphere
Human: “The red triangle is right-of the blue sphere” Word – Category
Mapping
10
Word – Category Mapping
Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Prep Learning
Action Knowledge
Procedural Memory
Preposition – Spatial RelMapping
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 4Preposition
Learning
- Lookup the category of red
from smem
Working Memory
Noun Learning
red color
Prep Learning
Obj1- bl3- sp7
Obj2-re23-tr17
Obj3- bl3- sq8
Prep. Learningright-of
red, triangle
blue, sphere
Human: “The red triangle is right-of the blue sphere”
Indexing
11
Noun/Adjective – Perceptual Symbol
Mapping
Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Prep Learning
Action Knowledge
Procedural Memory
Preposition – Spatial RelMapping
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 4Preposition
Learning
- Lookup the symbol for red + color from smem
Working Memory red color
Noun/Adjective – Perceptual Symbol Map
re23Obj1- bl3- sp7
Obj2-re23-tr17
Obj3- bl3- sq8
Prep. Learningright-of
red, triangle
blue, sphere
re23, triangle
Human: “The red triangle is right-of the blue sphere”
Indexing
Word – Category Mapping
Word – Category Mapping
12
Noun/Adjective – Perceptual Symbol
Mapping
Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Prep Learning
Action Knowledge
Procedural Memory
Preposition – Spatial RelMapping
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 4Preposition
Learning
- Update the descriptions with
perceptual symbols
Working Memory
Noun/Adjective – Perceptual Symbol Map
Obj1- bl3- sp7
Obj2-re23-tr17
Obj3- bl3- sq8
Prep. Learningright-of
red, triangle
blue, sphere
re23, trianglere23, tr17
bl3, spherebl3, sp7
Human: “The red triangle is right-of the blue sphere”
Indexing
Word – Category Mapping
13
Noun/Adjective – Perceptual Symbol
Mapping
Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Prep Learning
Action Knowledge
Procedural Memory
Preposition – Spatial RelMapping
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 4Preposition
Learning
- Look for the objects matching the description
Working MemoryObj1- bl3- sp7
Obj2-re23-tr17
Obj3- bl3- sq8
Prep. Learningright-of
re23, tr17
bl3, sp7
Obj2
Obj1
Human: “The red triangle is right-of the blue sphere”
Indexing
Word – Category Mapping
14
Obj2
Obj1
Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Action Knowledge
Procedural Memory
Preposition – Spatial RelMapping
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 4Preposition
Learning
- SVS updates model of objects in the world from the perception input.
Human: “The red triangle is right-of the blue sphere”
Working MemoryObj1- bl3- sp7
Obj2-re23-tr17
Obj3- bl3- sq8
Prep. Learningright-of
Obj1Obj2Obj3Spatial Primitives
Prep Learning
Word – Category Mapping
15
Obj2
Obj1
Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Action Knowledge
Procedural Memory
Preposition – Spatial RelMapping
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 4Preposition
Learning
- SVS is queried for primitive
spatial relationships
between Obj1 and Obj2
Human: “The red triangle is right-of the blue sphere”
Prep Learning
Working MemoryObj1- bl3- sp7
Obj2-re23-tr17
Obj3- bl3- sq8
Prep. Learningright-of
Obj1
Obj2
Obj3
Obj2
Obj1?
Spatial Primitives
Word – Category Mapping
16Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Action Knowledge
Procedural Memory
Preposition – Spatial RelMapping
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 4Preposition
Learning
- The results of the query are
stored in working memory.
Human: “The red triangle is right-of the blue sphere”
Prep Learning
y -alignedz-aligned+x
Working Memoryright-of
Obj1- bl3- sp7
Obj2-re23-tr17
Obj3- bl3- sq8
Prep. Learning
Word – Category Mapping
17Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Action Knowledge
Procedural Memory
Preposition – Spatial RelMapping
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 4Preposition
Learning
- List of spatial primitives are
abstracted into general spatial
relationship and mapped to “right-
of.”
Human: “The red triangle is right-of the blue sphere”
Prep Learning
Working MemoryObj1- bl3- sp7
Obj2-re23-tr17
Obj3- bl3- sq8
SR11
y-aligned z-aligned +xright-of
Prep. Learning
Word – Category Mapping
18
Preposition – Spatial RelMapping
Preposition – Spatial RelMapping
Perception
LGSoar
Interaction
Indexing
Verb Learning
Noun Learning
Action Knowledge
Procedural Memory
Verb – Operator Mapping
Noun/Adjective – Perceptual Symbol
Mapping
Semantic Memory
Primitive Verb – Operator Mapping
Episodic Memory
Agent’s Experiences
Spatial Visual System
Spatial PrimitivesAction
Phase 4Preposition
Learning
- Learned spatial relationship mapping to “right-of” is
stored in semantic memory.
Human: “The red triangle is right-of the blue sphere”
Prep Learning
Working MemoryObj1- bl3- sp7
Obj2-re23-tr17
Obj3- bl3- sq8
SR11
y-aligned z-aligned +xright-of
Prep. Learning
Word – Category Mapping
19
Refining learned knowledge
Is this triangle also to the right of the square?
Disjunctive primitive relationships in training are removed.
20
Learning is interactive
• Learning can be initiated by instructor or agent
“What is behind the blue sphere?”- “I don’t know the preposition behind. Please teach me with
examples.”“The green square is behind the red triangle.” “Finished teaching.”- “The blue square is behind the blue sphere.”
22
Current Uses
• Describing the scene• Specifying objects in the scene• Tracking objects and relationships– “The goal is that the red triangle is inside the
pantry”• Projecting relationships back to a point in 3D
space to perform actions
23
Evaluation
Nuggets• Learns very fast• Works well for simple relationships and projection/tracking
Coals• Cannot learn complex relationships with many objects• Cannot learn certain types of relationships such as diagonal• Not robust to mistakes in training examples
24
Future Work
• Complex multi-object relationships– For example in a line or a circle
• More complex spatial relationships– Using distance/contacting etc.– Pattern matching (diagonal)– Will require more training examples
• Projection with multiple relationships– “put this inside the pantry and to the right of the blue object”