Grant: Discovery Launch Supplement
University of Alberta
$12,500.00
Modeling Embodied Dexterity for Open-ended Robot Learning
FY 2024-25
Summary
Modeling Embodied Dexterity for Open-ended Robot Learning
Despite the rapid progress in artificial intelligence, machine learning, and many sub-fields of robotics, the progress in robot dexterity has been slow. While recent works including hierarchical reinforcement learning and option/skill discovery have been proposed, two core problems have not yet been well understood, (1) what computational model is sufficient to encode complex dynamics of dexterous manipulation skills, and (2) how to make the model self-adaptive, expandable and extendable to the changes of environments, task objectives and even sensory modalities--- the open-ended robot learning setting.
Our approach of tackling the open-ended robot learning challenge is to start with the robot's body control---how motor skills are modelled and how they are generally reused in downstream task learning. Our unique perspectives are (1) shifting the classic dynamic system based approach of modelling motor skills to a predictive representation learning approach will bring better generality with modern learning methods, and (2) shifting the world model approach of building the backbones for open-ended robot learning to an embedded agent model construction approach will bring feasibility in open-ended learning settings.
Our long-term objective is to discover general methods (learning architecture and algorithms) to model embodied robot dexterous manipulation skills that are self-adaptive, and extendable to robot task learning in open-ended environments. We break it down into three research project aims:
Aim 1: Modelling Embodied Dexterity with Predictive Representations.
Aim 2: Extending Embodied Dexterity to Contact-rich Manipulation Skills.
Aim 3: Embodied Dexterity for Open-ended Robot Learning.
Impacts
In Robotics, this research investigates novel learning architectures and mechanisms that can bootstrap robot learning in open-ended environments with embodied motor skills. The study of embodied dexterity will facilitate the creating of robots with human-like manipulation skills that can be deployed with greater generality, and thus will have the potential to accelerate the adoption of robots in senior care, and assistive devices for handicapped people. The University of Alberta has a strong academic profile in assistive robotics and rehabilitation medicine research. By collaborating with the researchers, the application potential of the proposed research will create significant societal impact.
In Machine Learning, this research investigates a possible approach to building embodied intelligence. Compared to the prevalent approach that relies on a huge amount of data and computation resources, our approach gradually refines the learned representation from the agent's past experience and updates the learned representation/model through agent-environment interactions. This research will shed light on the long-standing challenge in artificial intelligence and robotics - Moravec's paradox.
Program
Discovery Launch Supplement
Recipient
Location
Alberta, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data
As published
Application id
DGECR-2024-00327
Name nom
Jin, Jun
Department département
Electrical and Computer Engineering
Organization id
9
Institution établissement
University of Alberta
Province en
Alberta
Province fr
Alberta
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2024
Competition year année de concours
2024
Award amount
12500
Program id
DGECR
Program name en
Discovery Launch Supplement
Program name fr
Supplément Tremplin vers la découverte
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1507
Committee name en
Computer Science
Committee name fr
Sciences informatiques
Area of application code
1206
Area of application group en
Advancement of knowledge
Area of application group fr
Avancement des connaissances
Area of application en
Engineering
Area of application fr
Génie
Research subject code
2600
Research subject group en
Robotics
Research subject group fr
Robotique
Research subject en
Robotics
Research subject fr
Robotique
Field of research list names en
Not available
Field of research list names fr
Non disponible
Socioeconomic objective list names en
Not available
Socioeconomic objective list names fr
Non disponible
Application title
Modeling Embodied Dexterity for Open-ended Robot Learning
Keywords
robot learning; reinforcement learning; representation learning; motor skill learning for robotics; continual learning; embodied intelligence; sensorimotor abstraction
Application summary
Modeling Embodied Dexterity for Open-ended Robot Learning
Despite the rapid progress in artificial intelligence, machine learning, and many sub-fields of robotics, the progress in robot dexterity has been slow. While recent works including hierarchical reinforcement learning and option/skill discovery have been proposed, two core problems have not yet been well understood, (1) what computational model is sufficient to encode complex dynamics of dexterous manipulation skills, and (2) how to make the model self-adaptive, expandable and extendable to the changes of environments, task objectives and even sensory modalities--- the open-ended robot learning setting.
Our approach of tackling the open-ended robot learning challenge is to start with the robot's body control---how motor skills are modelled and how they are generally reused in downstream task learning. Our unique perspectives are (1) shifting the classic dynamic system based approach of modelling motor skills to a predictive representation learning approach will bring better generality with modern learning methods, and (2) shifting the world model approach of building the backbones for open-ended robot learning to an embedded agent model construction approach will bring feasibility in open-ended learning settings.
Our long-term objective is to discover general methods (learning architecture and algorithms) to model embodied robot dexterous manipulation skills that are self-adaptive, and extendable to robot task learning in open-ended environments. We break it down into three research project aims:
Aim 1: Modelling Embodied Dexterity with Predictive Representations.
Aim 2: Extending Embodied Dexterity to Contact-rich Manipulation Skills.
Aim 3: Embodied Dexterity for Open-ended Robot Learning.
Impacts
In Robotics, this research investigates novel learning architectures and mechanisms that can bootstrap robot learning in open-ended environments with embodied motor skills. The study of embodied dexterity will facilitate the creating of robots with human-like manipulation skills that can be deployed with greater generality, and thus will have the potential to accelerate the adoption of robots in senior care, and assistive devices for handicapped people. The University of Alberta has a strong academic profile in assistive robotics and rehabilitation medicine research. By collaborating with the researchers, the application potential of the proposed research will create significant societal impact.
In Machine Learning, this research investigates a possible approach to building embodied intelligence. Compared to the prevalent approach that relies on a huge amount of data and computation resources, our approach gradually refines the learned representation from the agent's past experience and updates the learned representation/model through agent-environment interactions. This research will shed light on the long-standing challenge in artificial intelligence and robotics - Moravec's paradox.