Within this repository we are attempting to achieve computational perception via Reinforcement learning and robots.
There are several parts to this and thus, several layers of documentation.
- Provides steps required to install ROS, and begin to send command to a dynamixel servo, as well as reading data from the servo.
- Describes answering 3 predictive questions about the future using 3 different general value functions.
- Presents an architecture that is capable of learning thousands of general value functions in parallel, based of just the sensorimotor information from the servos.
- Demonstrates how to alter the behavior of the robot based on these predictions
- Demonstrates how to use error measures that reflect learning progress.
- Presents an example of using policy gradient to control the actuator directly be adjusting the policy rather than by adjusting the value function.
- Demonstrates using Actor Critic methods
- Demonstrates using continuous actions - by parameterizing the mean and variance by which actions are taken.