Reinforcement Learning-Based Controllers for
Flapping-wing Microrobots
Name
singh-meenu-meng-eecs-2026-thesis.pdf
Size
4.57 MB
Format
Adobe PDF
Checksum (MD5)
756581993a5aa943edd06cd202e246d6
Author(s)
Singh, Meenakshi
Advisor(s)
Chen, YuFeng
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
Micro-scale soft aerial robots face limitations due to their extremely small scale that are often unaccounted for by traditional control approaches for aerial vehicles such as quadrotors. The environmental disturbances and system measurement errors are often larger relative to the size of the robot, which leads to magnifications in trajectory error. The physical hardware limitations also mean that controllers must be computationally efficient. In this work, we investigate the use of imitation learning combined with adversarial reward learning to enable robust flight control in a microrobotic flapping-wing platform, SoftFly. We generate expert demonstrations through behavioral cloning in both standard and domain-randomized environments, and evaluate the effectiveness of fine-tuning policies with PPO and AIRL. Our experiments demonstrate that behavioral cloning provides a strong initial policy, and AIRL initialized from BC can improve robustness and reduce positional errors even with limited expert data.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
