A Simple Example for Imitation Learning with Dataset Aggregation (DAGGER) on Torcs Env
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
.gitignore
README.md
autostart.sh
dagger.py
gym_torcs.py
snakeoil3_gym.py

README.md

Imitation Learning with Dataset Aggregation (DAGGER) on Torcs Env

This repository implements a simple algorithm for imitation learning: DAGGER. In this example, the agent only learns to control the steer [-1, 1], the speed is computed automatically in gym_torcs.TorcsEnv.

Requirements

  1. Ubuntu (I only test on this)
  2. Python 3
  3. TensorLayer and TensorFlow
  4. Gym-Torcs

Setting Up

It is a little bit boring to set up the environment, but any incorrect configurations will lead to FAILURE. After installing Gym-Torcs, please follow the instructions to confirm everything work well:

  • Open a terminal:

    • Run sudo torcs -vision to start a game
    • Race --> Practice --> Configure Race: set the driver to scr_server 1 instead of player
    • Open Torcs server by selecting Race --> Practice --> New Race: This should result that Torcs keeps a blue screen with several text information.
  • Open another terminal:

    • Run python snakeoil3_gym.py on another terminal, it will shows how the fake AI control the car.
    • Press F2 to see the driver view.
  • Set image size to 64x64x3:

    • The model is trained on 64x64 RGB observation.
    • Run sudo torcs -vision to start a game
    • Options --> Display --> select 64x64 --> Apply

Usage

Make sure everything above work well and then run:

  • python dagger.py

It will start a Torcs server at the beginning of every episode, and terminate the server when the car crashs or the speed is too low. Note that, the self-contained gym_torcs.py is modified from Gym-Torcs, you can try different settings (like default speed, terminated speed) by modifying it.

Results

After Episode 1, the car crashes after 315 steps.

After Episode 3, the car does not crash anymore !!!

The number of steps and episodes might vary depending on the parameters initialization.

ENJOY !