Model-Based RL Multi-Tasking with ReLAx
This repository contains an implementation of multi-tasking with CEM actor.
Multi-Tasking concept:
CEM is a derivative-free optimizer which searches for optimal actions by iteratively generating random action sequences and evaluating them with fitted observation and reward models. First action from elite (by total rewards sum) sequence is then executed.
By replacing rewards calculated by fitted reward model with some user defined function we can alter RL agent's behavior without the need of retraining any model. Thus, achieving multi-task agent.
Default Behavior:
default_reward.mp4
Stand Vertically:
vertical_body.mp4
Front-Flip: