manavkataria / nvidia-autopilot-tensorflow Goto Github PK

Compute Steering Angle for keeping a self-driving car in a lane, given a single camera image as input. Read more at Udacity Challenge 2 Blog.

Deadline: October 28, 2016

A TensorFlow implementation of this Nvidia paper with some changes.

Use python train.py to train the model

Use python run.py to run the model on a live webcam feed

Use python run_dataset.py to run the model on the dataset

Download a Udacity Dataset and extract into a (new) $REPO_ROOT/dataset folder

Use Udacity ROS Reader repository to extract the dataset using Docker containers.

Udacity Official: Simulator used for evaluation.

As of October 15:

• Started: October 10, 2016
• Got the environment up and running (on MacOS host)
• Successfully trained a NN using a open source implementation of nVidia-end-to-end-learning network (took ~9 hours)
• Upgraded to nVidia-Cuda Multi Core Hardware which brought training time down to ~1 hour
• Trained model using Udacity Sunny data from 09/29/2016 (12:40 mins)
  • RMSE was too low
  • Model was overfit
• WIP: Filter training to kill low speed and steep steering angles
• Improve Neural Network to overcome Observed Limitations
• Prepare Training Data with Scene Augmentation for drift recovery training
• Try Alternative ML methods if needed
• Move Dev Environment to Cloud (Google / Amazon)

Udacity Dataset Torrents