- First, we model the environment dynamics of an F1/10 self-driving car using a meural network.
- This is learned using sample trajectories generated from the simulator.
- The objective function calculates far the car is from the centerline of the track.
- Thus learned model predicts the latent state and the reward for the action taken at the current state.
- While real physics model is availble, it does not address the noise and idiosyncracies present in the environment. A neural network, on the other hand, can learn the idiosyncracies of the environment not factored in by an ideal physics model.
- Then, we learn motion plans using deep cross entropy (CEM). For this, we make use of the simulator instead of the learned state space model.
Following is a demo of the car using the policy generated using CEM to run on the racetrack.
This is a collborative work by Aashish Adhikari, Niraj Basnet, and Ashwin Vinoo.