A Novel Reinforcement Learning Method for Efficient Cross-Training Between Real and Simulated Robots
Publication Date
Summer 2024
Degree Type
Thesis
Degree Name
Master of Science (MS)
Department
Computer Engineering
Advisor
Stas Tiomkin; Harry Li; Mahima Agumbe Suresh
Abstract
Training reinforcement learning policy in a simulated environment can be a go-to choice for many research topics, as a simulated environment provides flexibility and costs less than building a physical environment for training. However, policy trained in a simulated environment often fails to transfer to the real environment for problems that have a more complex dynamics. As the solution, Sim2Real was proposed and it categorizes techniques that improve the performance of the transfer from simulation to reality. Sim2Real is an area of study that focuses on utilizing simulation data to train models that can apply to real world environment. It can greatly boost the range of reinforcement learning application by lowering the bar of training effective policy, furthermore it makes tasks that are dangerous or difficult to experiment in the real world accessible. This work will propose an approach of utilizing ensemble variance from the physical and simulated ensemble to bootstrap training with limited real world data, and it will be examined and validated by a dedicated cyber-physical system. Data collected by physical robots can enhance samples collected from simulated environment with a ensemble learning approach. Enhanced Simulated data can also provide feedback to real robots’ learning process. This learning cycle can reduce the cost of training without sacrificing the model’s robustness.
Recommended Citation
Liang, Zhonglin, "A Novel Reinforcement Learning Method for Efficient Cross-Training Between Real and Simulated Robots" (2024). Master's Theses. 5540.
DOI: https://doi.org/10.31979/etd.wp8f-99ag
https://scholarworks.sjsu.edu/etd_theses/5540
