Artificial agent learns representative trajectories by imitation

How does a cheetah learn to chase gazelles? By imitating their motion!

In this application:

• I introduced a method for trajectory imitation by bringing ideas from unsupervised image to image translation to the reinforcement learning realm.
• The agent is given the current position of the trajectory to imitate and, at each step, has to decide the steering direction and speed to guess the next position.
• By presenting the agent with many episodes involving a distribution of trajectories, it learns to rollout a typical motion from a single initial state.

For more details, check out the paper.