Grant: Discovery Grants Program - Individual
The University of British Columbia
$36,500.00
Enhancing the Concepts of Machine Intelligence in the Context of Mechanical Dynamic Systems
FY 2023-24
Summary
The proposed research will enhance machine intelligence as applied to mechanical systems such as robots, particularly by incorporating a dynamic model (physics) of the system. Three paramount issues of machine intelligence, which have shortcomings, will be explored, improved, and verified: 1. Characteristics of Intelligence, and the Incorporation of Self-awareness; 2. Incorporation of Physics into Machine Learning; 3. Domain Transformation. New methodologies will be developed and existing methodologies will be enhanced, and the developments will be validated in laboratory, for a robotic homecare application. Under self-awareness, the proposed work will consider both the machine and the required task simultaneously, focusing on the specific capabilities (what and at what level) that the machine possesses in order to perform the required task. Furthermore, while the existing work rely on massive amounts of input-output data to ascertain self-awareness, through approaches of "learning" without incorporating the known physics of the machine, the proposed work will incorporate a dynamic model of the machine, in addition to the conventionally available data. The resulting performance measure will determine the ability of a machine to carry out a task, and hence the "best" machine for the task. In the context of incorporating physics into machine learning through a convolutional neural network (CNN), the inputs to the CNN are the "conditions" of the task, and the outputs of the network are the corresponding "actions" by the machine. First, a range of conditions (that may be encountered in the application domain) and the corresponding actions will be determined, from an analytical model of the machine. This model-based data, along with the conventionally available (labeled) data, will be used in training the CNN. Domain transformation is pertinent when a process is learned in one location or environment (the source domain) and the learned knowledge is implemented and practically utilized in another location of environment (the target domain). In the source domain the collected data is used to train neural network (NN). The trained NN would be implemented and practically utilized in the target domain. The output data (actions) corresponding to the input data (conditions) is known (i.e., labeled) in the training data. The output data corresponding to the input data is not known a priori (i.e., unlabeled) in the target domain. Adapting the source domain data to the target domain is key in domain transformation. A physics-aware domain Adaptation (PADA) learning scheme is developed to determine the action policies from the simulated conditions of source environments, using the physics learned from the trajectories, while minimizing the errors due to domain transformation. The schemes before and after enhancement will be implemented on a robotic homecare system in laboratory, and the performance will be evaluated and compared.
Program
Discovery Grants Program - Individual
Recipient
Location
British Columbia, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data
As published
Application id
RGPIN-2023-03243
Name nom
de Silva, Clarence
Department département
Mechanical Engineering
Organization id
2
Institution établissement
The University of British Columbia
Province en
British Columbia
Province fr
Colombie-Britannique
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2023
Award amount
36500
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1512
Committee name en
Mechanical Engineering
Committee name fr
Génie mécanique
Area of application code
518
Area of application group en
Manufacturing processes and products
Area of application group fr
Procédés et produits industriels
Area of application en
Medical equipment and apparatus
Area of application fr
Appareils et instruments médicaux
Research subject code
2109
Research subject group en
Mechanical engineering
Research subject group fr
Mechanical systems and instrumentation
Research subject en
Génie mécanique
Research subject fr
Instrumentation et systèmes mécaniques
Application title
Enhancing the Concepts of Machine Intelligence in the Context of Mechanical Dynamic Systems
Keyword
Characteristics of intelligence; Convolutional neural networks; Domain transformation; Dynamic models; Engineering process automation; Homecare robots; Machanical dynamic systems; Machine learning; Robts; Self-awareness
Application summary
The proposed research will enhance machine intelligence as applied to mechanical systems such as robots, particularly by incorporating a dynamic model (physics) of the system. Three paramount issues of machine intelligence, which have shortcomings, will be explored, improved, and verified: 1. Characteristics of Intelligence, and the Incorporation of Self-awareness; 2. Incorporation of Physics into Machine Learning; 3. Domain Transformation. New methodologies will be developed and existing methodologies will be enhanced, and the developments will be validated in laboratory, for a robotic homecare application. Under self-awareness, the proposed work will consider both the machine and the required task simultaneously, focusing on the specific capabilities (what and at what level) that the machine possesses in order to perform the required task. Furthermore, while the existing work rely on massive amounts of input-output data to ascertain self-awareness, through approaches of "learning" without incorporating the known physics of the machine, the proposed work will incorporate a dynamic model of the machine, in addition to the conventionally available data. The resulting performance measure will determine the ability of a machine to carry out a task, and hence the "best" machine for the task. In the context of incorporating physics into machine learning through a convolutional neural network (CNN), the inputs to the CNN are the "conditions" of the task, and the outputs of the network are the corresponding "actions" by the machine. First, a range of conditions (that may be encountered in the application domain) and the corresponding actions will be determined, from an analytical model of the machine. This model-based data, along with the conventionally available (labeled) data, will be used in training the CNN. Domain transformation is pertinent when a process is learned in one location or environment (the source domain) and the learned knowledge is implemented and practically utilized in another location of environment (the target domain). In the source domain the collected data is used to train neural network (NN). The trained NN would be implemented and practically utilized in the target domain. The output data (actions) corresponding to the input data (conditions) is known (i.e., labeled) in the training data. The output data corresponding to the input data is not known a priori (i.e., unlabeled) in the target domain. Adapting the source domain data to the target domain is key in domain transformation. A physics-aware domain Adaptation (PADA) learning scheme is developed to determine the action policies from the simulated conditions of source environments, using the physics learned from the trajectories, while minimizing the errors due to domain transformation. The schemes before and after enhancement will be implemented on a robotic homecare system in laboratory, and the performance will be evaluated and compared.