RL Based Slosh Control

Summary :

• Started by tuning hyperparameters of STC controller using RL in numerical simulations, surpassed the SOTA results
• Secured a grant from Artpark @ IISc Bangalore to realize the research on hardware
• Developed a Hardware AGV prototype with following features
  • 4 wheel holonomic meccanum drive
  • Localisation using fusion of 2 perpendicular single point lidars and an IMU
  • PID based path tracking
  • Fuzzy based velocity controller
  • Capacitance based slosh measurement
• Modelled the hardware AGV prototype using simple pendulum analogy
  • Estimated the parameters using translational excitation and quick stop strategy
• Used the estimated model implement 2nd order Sliding mode controller using Super Twisting Algorithm on the hardware.
• Used the estimated model to train RL agent to optimise STC parameters in simulation and applied the learned parameters to the hardware
• Used the estimated model to numerically train DDPG based model-free controller in simulation
• Applied the trained model-free controller on the hardware and fine-tuned its hyperparameters on hardware for some episodes
• Received the Singhal’s Tech for society award for most innovative, impactful and commercially viable bachelor’s thesis with a cash prize of 50k

Here you can find a video(mid-term review) of us explaining the task

Also, for more details, please check out the final thesis