Grant: Research Tools and Instruments
McGill University
$61,661.00
Safe Learning for Bipedal Robot Locomotion
EF 2023-24
Résumé
Machine learning for bipedal walking robots is known to be a complicated problem; either the algorithm does not converge into a solution, or it takes an extremely long time to train. However, most of this work is only done in simulation, and the results fail on the real system due to the discrepancy between the simulator and the hardware. Moreover, legged robots are often mechanically complex and hard to maintain without significant costs. The research of my group centers around optimal control and machine learning for robot motion. Recently, we have published two papers on using a data-driven approach for robot motion planning with stability guarantee. In optimal control, the stability of a system can be verified through the Lyapunov stability theory. Our previous work designs neural network architecture such that a model learned through deep learning intrinsically fulfills the Lyapunov stability theory, which certifies the safety of using data-driven methods on robotic hardware. While our former work was validated on robot manipulators, it opens several extremely exciting new directions for legged robots, from both hardware and software perspectives. The new research directions of my group can be categorized into the following themes: (1) learning a neural network for bipedal robots where guarantees the stability of the learned model, (2) refine a neural network on bipedal robots while ensuring the process is safe, and (3) using soft and bio-degradable materials for our bipedal robots. The success of this research program will significantly change the field of bipedal locomotion by (1) developing a framework for learning safe locomotion policy before deployment on a robotic system to reduce the risk the breaking the hardware, (2) closing the sim-to-real gaps so that researchers worldwide can focus on developing novel planning algorithms, and (3) using soft and bio-degradable materials to reduce the maintance cost of robotic research. Open Dynamics Robot Initiative (ODRI) has published a white paper on building open-source, low-cost, lightweight, and low-complexity robots, where one of the aims is to make a common benchmark for legged robot research. The design is highly modular; specifically, the users can build from basic components (such as an actuator) and assemble them into various legged systems. We seek innovations in the theory, practice and deployment of robotic intelligence on legged robots, whereby building our own robots designed by the Open Dynamic Robot Initiative stands out as the most cost-effective way of enabling experimentation in all the target environments while being open in design and software to enable flexible use by my students and collaborators.
Programme
Research Tools and Instruments
Bénéficiaire
Emplacement
Quebec, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data
Tel que publié
Application id
RTI-2024-00668
Name nom
Lin, Hsiu-Chin
Department département
Computer Science
Organization id
61
Institution établissement
McGill University
Province en
Québec
Province fr
Québec
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2024
Award amount
61661
Program id
RTI
Program name en
Research Tools and Instruments
Program name fr
Outils et instruments de recherche
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1607
Committee name en
Computer, Mathematical and Statistical Sciences (RTI)
Committee name fr
Sciences informatiques, mathématiques et statistiques (OIR)
Area of application code
801
Area of application group en
Information and communication services
Area of application group fr
Services d'information et de télécommunications
Area of application en
Computer software
Area of application fr
Logiciel
Research subject code
9999
Research subject group en
Not available
Research subject group fr
Not available
Research subject en
Non disponible
Research subject fr
Non disponible
Application title
Safe Learning for Bipedal Robot Locomotion
Keyword
Machine Learning; Motion Planning; Robotics; Safe Learning
Application summary
Machine learning for bipedal walking robots is known to be a complicated problem; either the algorithm does not converge into a solution, or it takes an extremely long time to train. However, most of this work is only done in simulation, and the results fail on the real system due to the discrepancy between the simulator and the hardware. Moreover, legged robots are often mechanically complex and hard to maintain without significant costs. The research of my group centers around optimal control and machine learning for robot motion. Recently, we have published two papers on using a data-driven approach for robot motion planning with stability guarantee. In optimal control, the stability of a system can be verified through the Lyapunov stability theory. Our previous work designs neural network architecture such that a model learned through deep learning intrinsically fulfills the Lyapunov stability theory, which certifies the safety of using data-driven methods on robotic hardware. While our former work was validated on robot manipulators, it opens several extremely exciting new directions for legged robots, from both hardware and software perspectives. The new research directions of my group can be categorized into the following themes: (1) learning a neural network for bipedal robots where guarantees the stability of the learned model, (2) refine a neural network on bipedal robots while ensuring the process is safe, and (3) using soft and bio-degradable materials for our bipedal robots. The success of this research program will significantly change the field of bipedal locomotion by (1) developing a framework for learning safe locomotion policy before deployment on a robotic system to reduce the risk the breaking the hardware, (2) closing the sim-to-real gaps so that researchers worldwide can focus on developing novel planning algorithms, and (3) using soft and bio-degradable materials to reduce the maintance cost of robotic research. Open Dynamics Robot Initiative (ODRI) has published a white paper on building open-source, low-cost, lightweight, and low-complexity robots, where one of the aims is to make a common benchmark for legged robot research. The design is highly modular; specifically, the users can build from basic components (such as an actuator) and assemble them into various legged systems. We seek innovations in the theory, practice and deployment of robotic intelligence on legged robots, whereby building our own robots designed by the Open Dynamic Robot Initiative stands out as the most cost-effective way of enabling experimentation in all the target environments while being open in design and software to enable flexible use by my students and collaborators.