Chapter 10. The Explorer System in Cognitive Systems, Christensen et al. Course: Robots Learning from Humans On, Kyoung-Woon Biointelligence Laboratory.

Chapter 10. The Explorer Systemin Cognitive Systems, Christensen et al.

Course: Robots Learning from Humans

On, Kyoung-Woon

Biointelligence Laboratory

School of Computer Science and Engineering

Seoul National University

http://bi.snu.ac.kr

Contents Introduction

System Overview Navigation SA Object SA Place SA Conceptual Mapping SA

Spatial Modeling and Reasoning Map Acquisition Acquiring the Conceptual Map Cross-Modal Spatial Knowledge Sharing

Planning

Scenario: Find Object

Conclusion

Introduction

Chapter 10. The Explore System

Introduction PROBLEM: Modeling space, Acting in this space and Reasoning about it.

SETTING: Robot moving around in an initially Unknown, Large scale, environment Inhabited by humans.

MOTIVATION: To study the problems that occur when an intelligent robot must interact with humans in a rich and complex environment.

Introduction Construction of Spatial models from sensor data

Simultaneous Localization And Mapping

Introduction Spatial model: Sensors input + Human input

Two main modes : Robot explores space together with a user in a home tour fashion. Human Augmented Map-

ping paradigm. Fully autonomous exploration where the robot moves with the purpose of covering space.

Ability to perform tasks autonomously : Performing a particular task to perfection << Acting within a flexible framework

Investigate two problems within this context: What information must be exchanged by different parts of the system to make this possible How the current state of the world should be represented during such exchanges.

System Overview

Navigation SA The metric map (produced by the SLAM) The navigation graph (NavGraphProcess) Topological map (NavGraphProcess) Detecting and Tracking People Motion Control

Object SA The view planning component creates a plan for

which nodes to visit. The visual search can be performed using a pan-

tilt-zoom camera where an attention mechanism gradually guides the robot to zoom in closer.

System Overview

Place SA Assign nodes and areas to one of predefined seman-

tic place categories (e.g. an office, a corridor etc.) The results for both sensors are integrated by a cue

integration component, which provides beliefs about a place category for the current viewpoint.

The area categorization provides important infor-mation when reasoning about space.

Conceptual Mapping SA Maintains a symbolic representation of space suit-

able for situated action and interaction. It represents spatial areas, objects in the environ-

ment, and abstract properties of persons.

Spatial Modeling and Reason-ing

Spatial Modeling and Reasoning : Map Acquisition

Human Augmented Mapping Guided tour scenario. User walks up to the robot and initiates the mapping process with a com-

mand like “follow me!”. The robot continuously tracks the position of the user and follows them through the environment. As the robot moves, the spatial model is built from the sensor data.

Interaction is not master/slave-like. User can query the spatial knowledge of the robot throughout the HAM

sessions.

Spatial Modeling and Reasoning : Map Acquisition

Autonomous Exploration Explorer uses a frontier-based strategy [18] for autonomous exploration:

FREE, OCCUPIED, UNKNOWN (Frontier: boarder between FREE & UNKNOWN space)

Exploration is considered complete when there are no more reachable fron-tiers.

Because the way the navigation graph is built requires the robot to move, exploration is not only about having the sensor see all parts of the room, but the robot needs to move there as well

Spatial Modeling and Reasoning :Acquiring the Conceptual Map

System comes with a rich conceptual ontology taxonomies of indoor area types, of commonly found objects, and of

different spatio-topological relations between area instances and ob-ject instances.

the system is started with a blank map, the A-Box of the reasoner is empty.

it will contain area, person and object instances, and their relations, as contained in the loaded map.

Infer new or more specific knowledge based on partial in-formation By combining and reasoning over instances and their relations, the

reasoned can infer more specific concepts for those instances. Make the information inside the Conceptual Mapping SA

available to other sub-architectures.

Spatial Modeling and Reasoning :Cross-Modal Spatial Knowledge Sharing

Current Spatial Context A robot proxy representing the physical robot itself An area proxy for the area the robot is currently in A position relation connecting the robot proxy with its area A person proxy for each person currently being tracked by the people

tracking module area proxies for each person position relation proxies between the above An object proxy for each object belonging to an Area that is being repre-

sented position proxies connecting each object and its corresponding Area Close relation proxies between the robot and persons that are near to it

The Robot Proxy exactly one robot proxy on the binder. The only feature of this proxy is its Concept: robot.

Area Proxies The navigation subsystem divides space into areas, based on door nodes the sole binding feature AreaID

Object Proxies feature Concept, describing the particular class of object that it belongs to –

book or mug. The feature can be used both for binding and for executing a plan.

Person Proxies Concept: always set to person Location: last observed metric position PersonID: unique identifier

Position Relation Proxies Label: always set to position OtherSourceID: set to the ID of the navigation sub-architecture TemporalFrame: PERCEIVED for directly perceived entities; ASSERTED for

others

Closeness Relation Proxies Label: always set to close OtherSourceID: set to the ID of the navigation sub-architecture; negated TemporalFrame: always PERCEIVED

Planning

Chapter 10. The Explorer System in Cognitive Systems, Christensen et al. Course: Robots Learning from Humans On, Kyoung-Woon Biointelligence Laboratory.

snu cse biointelligence

modeling space

intelligent robot

human input

complex environment

object conclusion

robots learning

main modes

Documents

Woon en leven november

Cut-Based Graph Learning networks to Discover Compositional....

Hyun Kyoung Kim - Academic Commons

Woon-huurconstructies opvang breed: praktijkvoorbeelden

Woon Initiatief Nuenen (WIN)

'Accent Kook & Woon Eindejaarsfolder 2014

(C) 2002, SNU Biointelligence Lab, A Computer Scientist’s...

Bangsawan II Woon

Biointelligence Laboratory School of Computer Science and...

Competition, Preference Uncertainty, and Jamming: A...

Kim Jong Woon (KJW )

Seoul National University, Seoul 08826, South Korea AI...

Frank Chin Interview Koon Woon

Woon – en zorgbeleid

Yeon-Kyoung Chung

Illustrator Kyoung Hwan Kim