Optimizing the Swimming Pattern of a Robotic Fish Using Reinforcement Learning

Open Access
Beckman, Daniel
Area of Honors:
Mechanical Engineering
Bachelor of Science
Document Type:
Thesis Supervisors:
  • Bo Cheng, Thesis Supervisor
  • Hosam Kadry Fathy, Honors Advisor
  • reinforcement learning
  • robotics
  • fish locomotion
  • policy gradient
This thesis applies policy-gradient reinforcement learning to examine how thrust efficiency and swimming patterns of a robotic fish change with forward speed in a low Reynolds number environment. Its purpose is to provide knowledge about biological locomotive strategies and to inform the control of bio-inspired robotics. Existing research on fish swimming at high Reynolds number in the inertial regime shows that maximum efficiency occurs for Strouhal numbers between 0.25 and 0.35; however, comparatively little research exists for swimmers at low and intermediate Reynolds number where viscous effects play a role. This experiment provides quantitative data on thrust generation and efficiency for a flapping caudal fin at Reynolds number in the intermediate range of 101 to 102, corresponding to a transitional flow regime. It also determines optimal swimming patterns for swimmers in the transitional regime, which differ slightly from the patterns of inertial regime swimmers. Results of the experiment show that swimming at low Reynolds number requires high Strouhal numbers to generate enough thrust for steady-state motion. Additionally, thrust-to-power ratio, and therefore thrust efficiency, appears to increase with translational speed in the transitional regime.