Skip to content

Latest commit

 

History

History
57 lines (36 loc) · 2.59 KB

README.md

File metadata and controls

57 lines (36 loc) · 2.59 KB
Raw

Hand Gesture Controlled MountainCar-v0 (OpenAI-Gym)

A collaborative space to implement hand gesture control for OpenAI Gym's classic MountainCar-v0 environment using a Convolutional Neural Network (CNN).

Demo

MountainCar controll demo

Setup and Requirements

Create and activate a suitable conda environment named mnt_car:

conda env create -f environment.yaml
conda activate mnt_car

Execution

To properly control the MountainCar using hand gestures, ensure you run hand_gesture_control.py and mountain_car_play.py scripts at the same time. When gesture-controll.py starts recognizing your hand you can now run mountain-car-play.py. The game window will now be displayed. Click on the window and then start moving your index finger to left or right.

hand_gesture_control.py

This script initializes the ResNet9 model from utils_classes.py and loads our latest trained checkpoint. Once set in evaluation mode, it captures video frames in real-time and predicts hand gestures. Depending on the recognized gesture, a corresponding keyboard signal is sent:

  • Gesture Classes:
    • Class 0 (GO LEFT): sends a "left arrow" key signal.
    • Class 1 (DONT MOVE): sends a "down arrow" key signal.
    • Class 2 (GO RIGHT): sends a "right arrow" key signal.

The signals are intercepted by src/mountain_car_play.py to control the MountainCar game.

mountain_car_play.py

Allows you to play the MountainCar-v0 environment using the hand gestures recognized by the above script. When a recognized hand gesture corresponds to a control command (e.g., go left), the car in the environment will react accordingly.

Note: Make sure that both scripts are running simultaneously for the hand gesture control to work correctly with the MountainCar environment.

Training

To train your own gesture classifier run main.py and specify the training directory and all the other parameters in training.yaml:

n_batch: 64
epochs: 20
max_lr: 0.01
grad_clip: 0.1
weight_decay: 1e-4
opt_func: 'Adam'
train_path: 'C:\Users\henri\Documents\works\mountain_car\train_ARG'

Directories with different gestures should be present in the training directory following the name convention for each class: acc_l, acc_n and acc_r. The checkpoints are being saved in the checkpoints directory following the date and time. Please make sure to have the same number of training examples per class.

References