Skip to content

"Learning by Cheating" (CoRL 2019) submission for the 2020 CARLA Challenge

Notifications You must be signed in to change notification settings

bradyz/2020_CARLA_challenge

Repository files navigation

Learning by Cheating

teaser

Learning by Cheating
Dian Chen, Brady Zhou, Vladlen Koltun, Philipp Krähenbühl,
Conference on Robot Learning (CoRL 2019)
arXiv 1912.12294

If you find our repo to be useful in your research, please consider citing our work

@inproceedings{chen2019lbc
  author    = {Dian Chen and Brady Zhou and Vladlen Koltun and Philipp Kr\"ahenb\"uhl},
  title     = {Learning by Cheating},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2019},
}

The code in this repo is based off of link, which contains the code for the NoCrash and CoRL 17 benchmarks.

Installation

Clone this repo with all its submodules

git clone https://github.com/bradyz/2020_CARLA_challenge.git --recursive

All python packages used are specified in carla_project/requirements.txt.

This code uses CARLA 0.9.9 and works with CARLA 0.9.8, 0.9.10.1.

You will also need to install CARLA 0.9.10.1, along with the additional maps. See link for more instructions.

Dataset

We provide a dataset of over 70k samples collected over the 75 routes provided in leaderboard/data/routes_*.xml.

Link to full dataset (9 GB).

sample

The dataset is collected using leaderboard/team_code/autopilot.py, using painfully hand-designed rules (i.e. if pedestrian is 5 meters ahead, then brake).

Additionally, we change the weather for a single route once every couple of seconds to add visual diversity as a sort of on-the-fly augmentation. The simulator is run at 20 FPS, and we save the following data at 2 Hz.

  • Left, Center, and Right RGB Images at 256 x 144 resolution
  • A semantic segmentation rendered in the overhead view
  • World position and heading
  • Raw control (steer, throttle, brake)

Note: the overhead view does nothing to address obstructions, like overhead highways, etc.

We provide a sample trajectory in sample_data, which you can visualize by running

python3 -m carla_project.src.dataset sample_data/route_00/

Data Collection

The autopilot that we used to collect the data can use a lot of work and currently does not support stop signs.

If you're interested in recollecting data after changing the autopilot's driving behavior in leaderboard/team_code/autopilot.py, you can collect your own dataset by running the following.

First, spin up a CARLA server

./CarlaUE4.sh -quality-level=Epic -world-port=2000 -resx=800 -resy=600 -opengl

then run the agent.

export CARLA_ROOT=/home/bradyzhou/software/CARLA_0.9.10.1           # change to where you installed CARLA
export PORT=2000                                                    # change to port that CARLA is running on
export ROUTES=leaderboard/data/routes_training/route_19.xml         # change to desired route
export TEAM_AGENT=auto_pilot.py                                     # no need to change
export TEAM_CONFIG=sample_data                                      # change path to save data

./run_agent.sh

Run a pretrained model

Download the checkpoint from our Wandb project.

Navigate to one of the runs, like https://app.wandb.ai/bradyz/2020_carla_challenge_lbc/runs/command_coefficient=0.01_sample_by=even_stage2/files

Go to the "files" tab, and download the model weights, named "epoch=24.ckpt", and pass in the file path as the TEAM_CONFIG below.

Spin up a CARLA server

./CarlaUE4.sh -quality-level=Epic -world-port=2000 -resx=800 -resy=600 -opengl

then run the agent.

export CARLA_ROOT=/home/bradyzhou/software/CARLA_0.9.10.1           # change to where you installed CARLA
export PORT=2000                                                    # change to port that CARLA is running on
export ROUTES=leaderboard/data/routes_training/route_19.xml         # change to desired route
export TEAM_AGENT=image_agent.py                                    # no need to change
export TEAM_CONFIG=model.ckpt                                       # change path to checkpoint
export HAS_DISPLAY=1                                                # set to 0 if you don't want a debug window

./run_agent.sh

Training models from scratch

First, download and extract our provided dataset.

Then run the stage 1 training of the privileged agent.

python3 -m carla_project.src.map_model --dataset_dir /path/to/data --hack

We use wandb for logging, so navigate to the generated experiment page to visualize training.

Important: If you're interested in tuning hyperparameters, see carla_project/src/map_model.py for more detail.
To see what hyperparameters we used for our models, you can see all of them by navigating to the corresponding wandb run config.

sample

Training the sensorimotor agent (acts only on raw images) is similar, and can be done by

python3 -m carla_project.src.image_model --dataset_dir /path/to/data

Docker

Build the docker container to submit, make sure to edit scripts/Dockerfile.master appropriately.

sudo ./scripts/make_docker.sh

Spin up a CARLA server

./CarlaUE4.sh -quality-level=Epic -world-port=2000 -resx=800 -resy=600 -opengl

Now you can either run the docker container or run it interactively.

To run the docker container,

sudo docker run --net=host --gpus all -e NVIDIA_VISIBLE_DEVICES=0 -e REPETITIONS=1 -e DEBUG_CHALLENGE=0 -e PORT=2000 -e ROUTES=leaderboard/data/routes_devtest.xml -e CHECKPOINT_ENDPOINT=tmp.txt -e SCENARIOS=leaderboard/data/all_towns_traffic_scenarios_public.json leaderboard-user:latest ./leaderboard/scripts/run_evaluation.sh

Or if you need to debug something, you can run it interactively

sudo docker run --net=host --gpus all -it leaderboard-user:latest /bin/bash

Run the evaluation through the interactive shell.

export PORT=2000
export DEBUG_CHALLENGE=0
export REPETITIONS=1
export ROUTES=leaderboard/data/routes_devtest/route_00.xml         # change to desired route
export CHECKPOINT_ENDPOINT=tmp.txt
export SCENARIOS=leaderboard/data/all_towns_traffic_scenarios_public.json

conda activate python37

./leaderboard/scripts/run_evaluation.sh

About

"Learning by Cheating" (CoRL 2019) submission for the 2020 CARLA Challenge

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages