Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Learning Trough HRI
October 15th 2013
Department of Computer, Control and Management Engineering Sapienza, University of Rome
Guglielmo Gemignani
Outline
The Problem
Related Work
Semantic Mapping
Task Learning
Learning Trough HRI 1/19
Knowledgeable Robots
2/19
Gap between user expectations and robot functionality.
Difficulties arise from:
current capabilities of perception systems
difficulty of communicating with humans
ability to acquire, maintain and use knowledge
Robots are consumer products.
Learning Trough HRI
The General Idea
Symbiotic autonomy [1] and symbiotic robotics [2].
Exploit HRI to overcome the limitations of the robot.
1An effective personal mobile robot agent through symbiotic human-‐robot interaction. Rosenthal, Biswas and Veloso. 2Symbiotic robotic systems: Humans, robots, and smart environments. S. Coradeschi and A. Saffiotti.
Enable a robot to learn from humans in the same fashion a person might learn from another individual.
The Idea
3/19Learning Trough HRI
Useful Knowledge
Research Goal
Two types of knowledge:
Environmental and spatial knowledge (Semantic Mapping)
knowledge about actions to perform (Procedural Knowledge)
4/19Learning Trough HRI
Related Work
Research GoalSemantic Mapping
Automatic semantic mapping (Thrun et al., Saffiotti et al., Burgard et al., Mozos et al.)
Semantic mapping through HRI (Pronobis and Jensfelt, Kollar et al., Randelli and Bonanni et al.)
Task Learning
Robotic training through demonstration (Rybski and Voyles, Ng et al., Wang et al., Browning et al.)
Human/robot task training (Voyles et al., Kuniyoshi et al., Friedrich and Dillmann)
Task training through dialog (Lauria et al., Rybski et al., Meriçli et al.)
5/19Learning Trough HRI
Semantic Mapping
Research GoalFirst year contributed to three papers on semantic mapping [3] [4] [5]
3Knowledge representation for robots through human-‐robot interaction. Bastianelli, Bloisi, Capobianco, Gemignani, Iocchi and Nardi. 4On-‐line semantic mapping. Bastianelli, Bloisi, Capobianco, Cossu, Gemignani, Iocchi and Nardi.5Living with robots: Interactive environmental knowledge acquisition. Bastianelli, Bloisi, Capobianco, Gemignani, Iocchi and Nardi.
6/19Learning Trough HRI
Robot’s Environmental and Spatial Knowledge
Robot’s knowledge divided into two layers:
World Knowledge
Metric Map
Instance Signature Data Base
Cell Map
Topological Graph
Domain Knowledge
Conceptual Knowledge Base
Not to be interpreted as an extensional and intensional components of a classical knowledge base.
In fact, the world knowledge may be inconsistent with the domain knowledge, only used to support the robot when specific world knowledge is not available.
7/19Learning Trough HRI
Robot’s World Knowledge
From Metric Map To Semantic Map
8/19Learning Trough HRI
SapienzBot Video
9/19Learning Trough HRI
Future Work on Semantic Mapping
On-‐line mapping and area tagging
Human-‐robot interactions for knowledge revision and maintenance (e.g., dynamic position)
Perception and usage of the properties of the objects
Usage of one or multiple external conceptual knowledge bases
Sharing knowledge between robots and enabling them to communicate and query each other when in need.
10/19Learning Trough HRI
Procedural Knowledge Acquisition
Teach robots complex tasks based on previously known primitives actions
Rely on HRI in order to overcome the complexity of the task
11/19Learning Trough HRI
Procedural Knowledge Acquisition
Teach robots complex tasks based on previously known primitives actions
Rely on HRI in order to overcome the complexity of the task
How to approach the problem:
Define a formalism to describe primitive actions
Implement a basic approach for task learning through multi-‐modal human-‐robot interaction
Generalize and enhance the developed learning method
Refine the acquired knowledge through simulations and reinforcement learning
11/19Learning Trough HRI
Task LearningPetri Net Plans actions used to represent the primitives on which new commands are based.
“Check The Window” Command
12/19Learning Trough HRI
Task LearningPNP actions used to represent the primitives on which new commands are based.
Development of a dialog to enable the robot to start building complex actions:
How to distinguish an unknown command from a wrongly recognized phrase?
What are the HRIs that can help disambiguate this uncertainty?
13/19Learning Trough HRI
Task LearningPNP actions used to represent the primitives on which new commands are based.
Development of a dialog to enable the robot to start building complex actions:
How to distinguish an unknown command from a wrongly recognized phrase?
What are the HRIs that can help disambiguate this uncertainty?
User: Check the windowRobot: I don’t know the command “check the window”U: Ok, I’ll teach youR: I’m ready to learnU: First go in front of the windowR: Ok, then?U: Check if the window is open. If so guard it, else report hereR: Should I do something else afterwards?U: Wait for a stop message and a shut down messageR: anything else?U: End of plan.R: Plan learnt
13/19Learning Trough HRI
Task LearningPNP actions used to represent the primitives on which new commands are based.
Open issues:
What are the best mechanisms to concatenate primitives?
How can they be enhanced from a HRI point of view?
How can the lexicon usable by a user be expanded?
How the action grounding, managing and revision operations can be easily executed by a non-‐expert user?
Development of a dialog to enable the robot to start building complex actions:
How to distinguish an unknown command from a wrongly recognized phrase?
What are the HRIs that can help disambiguate this uncertainty?
14/19Learning Trough HRI
Reinforcement LearningWrongly thought actions as well as multiple grounding of the same commands should be taken into consideration.
15/19Learning Trough HRI
Reinforcement LearningWrongly thought actions as well as multiple grounding of the same commands should be taken into consideration.
Simulators
15/19Learning Trough HRI
Reinforcement LearningWrongly thought actions as well as multiple grounding of the same commands should be taken into consideration.
LearnPNP
Simulators
15/19Learning Trough HRI
Conclusion
Goal: Enable a robot to learn from humans in the same fashion a person might learn from another individual.
Semantic Mapping and Procedural Knowledge Acquisition
Related Work
Work Plan
Enhance the semantic mapping
Develop a novel task learning method based on HRI and PNP
Debug and refine task knowledge through simulators and LearnPNP
16/19Learning Trough HRI
Publications
E. Bastianelli, D. Bloisi, R. Capobianco, G. Gemignani, L. Iocchi, and D. Nardi, “Knowledge representation for robots through human-‐robot interaction,” in Proceedings of the Knowledge Representation and Reasoning in Robotics Workshop at ICLP 2013, 2013.
E. Bastianelli, D. Bloisi, R. Capobianco, F. Cossu, G. Gemignani, L. Iocchi, and D. Nardi, “On-‐line semantic mapping,” in Proceedings of the 16th International Conference on Advanced Robotics (ICRA), (in press), 2013.
E. Bastianelli, D. Bloisi, R. Capobianco, G. Gemignani, L. Iocchi, and D. Nardi, “Living with robots: Interactive environmental knowledge acquisition,” Artificial Intelligence Journal, (submitted), 2013.
17/19Learning Trough HRI
Activities
Conferences/WorkshopsRoboCup International Symposium6th International OberseminarKnowledge Representation and Reasoning in Robotics Workshop at ICLP 2013
PhD SchoolCITEC Summer School 2013
Other ActivitiesRoboCup Iran Open (1stPlace)RoboCup German Open (3rdPlace)RoboCup World Competition
18/19Learning Trough HRI
Name of the Course Professor Type Credits
A Systematic Analysis of Levels of Integration between Low-‐Level Reasoning and Task Planning
Peter Schüller C 0.5
Author Workshop Marco Schaerf, Massimo Mecella, Ralf Gestner, Elisa Magistrelli
C 0.5
Counterexample-‐guided Abstraction Refinement for Classical Planning
Jendrik Seipp C 0.5
Domain-‐Independent Planning at the service of Services Applications in Uncertain and Dynamic Domains
Eirini Kaldeli C 0.5
Getting the Most Out of Pattern Databases for Classical Planning Gabriele Röger C 0.5
Incremental LM-‐cut Florian Pommerening C 0.5
Modeling the Last Mile of the Smart Grid Andrea Pagani C 0.5
Protocols as Tangible Artifacts Farhad Arbab C 0.5
Seminar on Bayesian source separation in MEG Daniela Calvetti C 0.5
Seminar on Database Queries -‐ Logic and Complexity Moshe Y. Vardi C 0.5
Seminar on Hierarchical Bayesian Beamformers in Electroneurography
Erkki Somersalo C 0.5
Seminar on Reasoning About Strategies Aniello Murano C 0.5
Seminar on Turing and Artificial Intelligence Luigia Carlucci Aiello C 0.5
Stronger Abstraction Heuristics Through Perimeter Search Patrick Eyerich C 0.5
Towards Context Consistency in a Rule-‐Based Activity Recognition Architecture
Tuan Anh Nguyen C 0.5
Trial-‐based Heuristic Tree Search Thomas Keller C 0.5
Exams
18 CreditsTotal
Name of the Course Professor Type Credits Grade
CITEC Summer School 2013 Verified by Prof. Nardi B 2.5 30
Competition and Cooperation in Multi-‐agent Systems Prof. Luca Iocchi and Prof. Stefano Leonardi B 2.5 Very Good
Machine Learning Prof. Luca Iocchi B 2.5 30
Seminars in NLP Prof. Paola Velardi B 2.5 Excellent
19/19Learning Trough HRI
Petri Net Plans (PNPs)Primitives
12/15Learning Trough HRI
Petri Net Plans (PNPs)Primitives
Operators
12/15Learning Trough HRI