Awesome Open Source
Awesome Open Source

Build Status Documentation Status

Pylot is an autonomous vehicle platform for developing and testing autonomous vehicle components (e.g., perception, prediction, planning) on the CARLA simulator and real-world cars.

Setup instructions

Deploy using Docker

The easiest way to get Pylot running is to use our Docker image. Please ensure you have nvidia-docker on your machine before you start installing Pylot. In case you do not have nvidia-docker please run ./scripts/

We provide a Docker image with both Pylot and CARLA already setup.

docker pull erdosproject/pylot
nvidia-docker run -itd --name pylot -p 20022:22 erdosproject/pylot /bin/bash

Following, start the simulator in the container:

nvidia-docker exec -i -t pylot /home/erdos/workspace/pylot/scripts/

Finally, start Pylot in the container:

nvidia-docker exec -i -t pylot /bin/bash
cd workspace/pylot/
python3 --flagfile=configs/detection.conf

Visualizing components

In case you desire to visualize outputs of different components (e.g., bounding boxes), you have to forward X from the container. First, add your public ssh key to the ~/.ssh/authorized_keys in the container:

nvidia-docker cp ~/.ssh/ pylot:/home/erdos/.ssh/authorized_keys
nvidia-docker exec -i -t pylot sudo chown erdos /home/erdos/.ssh/authorized_keys
nvidia-docker exec -i -t pylot sudo service ssh start

Finally, ssh into the container with X forwarding:

ssh -p 20022 -X [email protected]
cd /home/erdos/workspace/pylot/
python3 --flagfile=configs/detection.conf --visualize_detected_obstacles

If everything worked ok, you should be able to see a visualization like the one below:

Pylot obstacle detection

Manual installation instructions

Alternatively, you can install Pylot on your base system by executing the following steps:

pip install -e ./

Next, start the simulator:

export CARLA_HOME=$PYLOT_HOME/dependencies/CARLA_0.9.10.1/

In a different terminal, setup the paths:

export CARLA_HOME=$PYLOT_HOME/dependencies/CARLA_0.9.10.1/
cd $PYLOT_HOME/scripts/
source ./

Finally, run Pylot:

python3 --flagfile=configs/detection.conf

Pylot components

Pylot comprises of several components: obstacle detection, traffic light detection, lane detection, obstacle tracking, localization, segmentation, fusion, prediction, planners, and control. Each component is implemented using one or more ERDOS operators and can be executed in isolation or with the entire Pylot application. Please read the Documentation for a more in depth description.

Pylot pipeline

Run the following command to see a demo of all the components, and the Pylot driving policy:

python3 --flagfile=configs/demo.conf

The demo will execute: obstacle detection, traffic light detection, segmentation, prediction, planning, and the driving policy.

You can also run components in isolation:

Obstacle detection

Pylot supports three object detection models: frcnn_resnet101, ssd-mobilenet-fpn-640 and ssdlite-mobilenet-v2. The following command runs a detector in isolation:

python3 --flagfile=configs/detection.conf

In case you want to evaluate the detector (i.e., compute mAP), you can run:

python3 --flagfile=configs/detection.conf --evaluate_obstacle_detection

In case you are not satisfied with the accuracy of our obstacle detector, you can run a perfect version of it:

python3 --flagfile=configs/perfect_detection.conf

If the detector does not run at your desired frequency, or if you want to track obstacles across frames, you can use a mix of detector plus tracker by running:

python3 --flagfile=configs/tracking.conf

Traffic light detection

Pylot has uses a separate component for traffic light detection and classification. The following command runs the component in isolation:

python3 --flagfile=configs/traffic_light.conf

In case you require higher accuracy, you can run perfect traffic light detection by passing the --perfect_traffic_light_detection flag.

Lane detection

python3 --flagfile=configs/lane_detection.conf

Obstacle tracking

python3 --flagfile=configs/tracking.conf


In order to run Pylot's segmentation component in isolation execute the following command:

python3 --flagfile=configs/segmentation.conf

Similarly, pass --perfect_segmentation if you desire ideal pixel semantic segmentation.


Pylot offers a simple linear prediction component:

python3 --flagfile=configs/prediction.conf


The planning component provides two planning options, which can be specified using the --planning_type flag:

  1. waypoint: a simple planner that follows predefined waypoints. These waypoints can either be either pre-specified or computed using the A-star planner part of the CARLA simulator map. The planner ensures that the ego-vehicle respects traffic lights, stops whenever there are obstacles in its path, but does not implement obstacle avoidance.
  2. frenet_optimal_trajectory: a Frenet Optimal Trajectory planner.
  3. rrt_star: a Rapidly-explory Random Tree planner.
  4. hybrid_astar: a Hybrid A* planner.
# To run the Frenet Optimal Trajectory planner.
python3 --flagfile=configs/frenet_optimal_trajectory_planner.conf

# To run the RRT* planner.
python3 --flagfile=configs/rrt_star_planner.conf

# To run the Hybrid A* planner.
python3 --flagfile=configs/hybrid_astar_planner.conf


Pylot supports three controllers, which can be specified using the control flag:

  1. pid: follows the waypoints computed by the planning component using a PID controller.
  2. mpc: uses model predictive control for speed and waypoint following.
  3. simulator_auto_pilot: uses the simulator auto pilot to drive on predefined routes. This controller drives independent of the output of the other components.

You can run all the components, together with one of the two policies by executing:

# Runs all components using the algorithms we implemented and the models we trained:
python3 --flagfile=configs/e2e.conf
# Runs the MPC
python3 --flagfile=configs/mpc.conf
# Runs the simulator auto pilot.
python3 --control=simulator_auto_pilot

Debug logs

In case you want to debug the application, you can active additional logging by passing: --log_file_name=pylot.log --v=1 to your command.

Data collection

Pylot also provides a script for collecting CARLA data such as: RGB images, segmented images, obstacle 2D bounding boxes, depth frames, point clouds, traffic lights, obstacle trajectories, and data in Chauffeur format.

Run python3 --help to see what data you can collect. Alternatively, you can inspect this configuration for an example of a data collection setup.

Build your own Docker image

In case you want to build your own images from the latest code, you can execute:

cd docker

The script creates two Docker images: one that contains the CARLA simulator and another one that contains ERDOS and Pylot.

CARLA autonomous driving challenge

Pylot can also be used as a baseline for executing on the CARLA Leaderboard routes. We provide an agent that offers reference implementations for perception (i.e., detection, tracking), localization (Extended Kalman filter), prediction, planning (e.g., waypoint follower, Frenet optimal trajectory, RRT*, Hybrid A*), and control.

To test this agent you can pull our image which has all the necessary software already installed.

docker pull erdosproject/pylot-carla-challenge
nvidia-docker run -itd --name pylot-challenge -p 20022:22 erdosproject/pylot-carla-challenge /bin/bash

Alternatively, you can manually install the dependencies on your machine by following the instructions provided below:

mkdir challenge
export CHALLENGE_ROOT=`pwd`
# Clone the challenge leaderboard repository.
git clone -b stable --single-branch
cd ${LEADERBOARD_ROOT} ; pip3 install -r requirements.txt ; cd ${CHALLENGE_ROOT}
# Clone the CARLA scenario runner repository. This is used by the leaderboard.
git clone -b leaderboard --single-branch
export SCENARIO_RUNNER_ROOT=${CHALLENGE_ROOT}/scenario_runner/
cd ${SCENARIO_RUNNER_ROOT} ; pip3 install -r requirements.txt ; cd ${CHALLENGE_ROOT}
# Checkout the CARLA challenge branch.
cd ${PYLOT_HOME} ; git checkout -b challenge origin/challenge
export CARLA_ROOT=Path to CARLA

Notes on the Pylot CARLA challenge agent

Similar to regular Pylot, the Challenge agent not only connects different reference implementation, but also provides the option of testing them in different configurations (e.g., test prediction, planning and control with perfect perception). This can be done by changing the flags in the challenge configuration according to the specification from the Pylot documentation.

Getting Involved

  • Community on Slack: Join our community on Slack for discussions about development, questions about usage, and feature requests.
  • Github Issues: For reporting bugs.

Get A Weekly Email With Trending Projects For These Topics
No Spam. Unsubscribe easily at any time.
python (53,198
machine-learning (3,567
autonomous-driving (79
control (77
prediction (68
autonomous-vehicles (59
planning (29
lane-detection (23
perception (18