CARLA Simulation based on RoadRunner

Key Features • Contents

Key Features

• Customized map on RoadRunner with high freedom
• Self built synchronous dataset from simulation
• Rewrite Tensorflow model with PyTorch for better computation velocity

Contents

• data_collection.py is to collect raw data from CARLA

  • Customized a map on RoadRunner and export it into Linux built CARLA.
  • Run make import to import map. (tutorial)
  • Set parameters like map/vehicle selection, vehicle motion params(velocity, etc), sensor params and number of datapoints to collect.
  • Run the data_collection.py to collect and store raw data into json/csv files.

• random_dataset.py is to build randomized synchronomous training dataset and test dataset from raw data

  • Function: Delete remains datapoints and some images to prepare for the dataset procession. Merge 9 time steps(raw data) into 1 data point.
  • Time steps every datapoint: 9
  • Data structure: ground truth/ action input.

• dataset_check.py is to check the collected raw data like velocity and generate corresponding image

• eval.py is the MPPI Control evaluation

  • Run CARLA to get sensor data
  • Then export these data into model to generate motion prediction.
  • With motion prediction and reward function, get an optimal path (angular velocity and linear velocity list)
  • Move vehicle in CARLA based on optimal path.
  • Get new sensor data, repeat steps above.

• model.py is for the CNN-LSTM model (from baseline project BADGR)

• trainer.py is for training

• real_time_plot.py is for real time data plot, like RGB image and lidar info.

  • Launch the CARLA simulator.
  • Adjust ego-vehicle into auto-driving mode.
  • Run real_time_plot.py to generate real time plot of multiple sensors.

Note: This project is actually unfinished. The evaluation part still has something wrong. But limited by my PC hardware and my PhD program isn't about mobile robot, this project stops here.