Following is a list of open, non-exhaustive questions that demand special care to deliver better reinforcement learning models in the field of robotics:
- How do we automate the process of state-action space representation?
- State-action spaces in robotics is continuous and multi-dimensional. The high-dimensionality and continuous nature of the state and action space makes the process of representation selection difficult to automate.
- State approximation is also an open question to deal with and is under intense study.
- How do we generate a reward function from the data received?
- The success of a reinforcement learning algorithm is highly dependent on the quality of the reward function, its coverage of different state representation, ...