Equipment: Cart-pole simulator and real cart-pole used by Tobi Delbruck for MPC tests

Possible directions:

• Online Spiking RL (R-Eprop) for non-ideal cartpole stabilization solving in simulation after transfer learning of weights for ideal cartpole trained with DQN 
• Deep Q-Learning for cartpole swing-up in simulation
• Implementation of inference DQN on real cartpole
• Possible simulation of cartpole with event-based tracked input instead of positions + velocities

Possible Challenges:

• Use an event-based tracking algorithm to compute the shape coordinates over time
• Define the RL policy (e.g., actions, state, reward) in simulation based on an event-camera output
• Transfer the policy to the real cart-pole
• Learn how to control the cart-pole