PlanT: Explainable Planning Transformers via Object-Level Representations

News:
19.01.2023: We released the code to generate the attention visualization.
02.12.2022: We released the perception checkpoint and the code for the SENSORS and MAP track agent. Conda environment needs to be updated. Checkpoints of the perception are in the checkpoint folder. Please download again.
11.11.2022: We made some changes in the agent files to ensure compatibility with our perception PlanT. We therefore uploaded new checkpoint files. The old one does not work anymore with the current code.

Project Page | Paper | Supplementary

This repository provides code for the following paper:

• Katrin Renz, Kashyap Chitta, Otniel-Bogdan Mercea, A. Sophia Koepke, Zeynep Akata and Andreas Geiger, PlanT: Explainable Planning Transformers via Object-Level Representations, CoRL 2022.

Content

• Setup
• Data and models
• Data generation
• Training
• Evaluation
• Explainability
• Perception PlanT
• Citation

Setup

First, you have to install carla and the conda environment.

# 1. Clone this repository
git clone https://github.com/autonomousvision/plant.git
cd plant
# 2. Setup Carla
# if you have carla already installed, skip the next step AND
# adapt the carla path in setup_env.sh before executing step 3.
chmod +x setup_carla.sh
./setup_carla.sh
# 3. Setup conda environment
chmod +x setup_env.sh
./setup_env.sh

conda activate plant
pip install -U openmim
mim install mmcv-full==1.7.0
pip install mmdet

Data and models

You can download our pretrained PlanT models by executing:

chmod +x download.sh
./download.sh

To download our 3x dataset run:

chmod +x download_data.sh
./download_data.sh

Data generation

You can download our dataset or generate your own dataset. In order to generate your own one you first need to start a Carla server:

# with display
./carla/CarlaUE4.sh --world-port=2000 -opengl

# without display
SDL_VIDEODRIVER=offscreen SDL_HINT_CUDA_DEVICE=0 ./carla/CarlaUE4.sh --world-port=2000 -opengl

To generate the data for the route specified in carla_agent_files/config/eval/train.yaml you can run

python leaderboard/scripts/run_evaluation.py user=$USER experiments=datagen eval=train

If you want to also save the sensor data that we used to train the perception module you can add the flag experiments.SAVE_SENSORS=1.

To generate the whole dataset you can use the datagen.sh file.

Training

To run the PlanT training on the 3x dataset, run:

python training/PlanT/lit_train.py user=$USER model=PlanT

To change any hyperparameters have a look at training/config/model/PlanT.yaml. For general training settings (e.g., #GPUs) check training/config/config.yaml.

Evaluation

This evaluates the PlanT model on the specified benchmark (default: longest6). The config is specified in the folder carla_agent_files/config.

Start a Carla server (see Data generation).
When the server is running, start the evaluation with:

python leaderboard/scripts/run_evaluation.py user=$USER experiments=PlanTmedium3x eval=longest6

You can find the results of the evaluation in a newly created evaluation folder inside the model folder. If you want to have a (very minimalistic) visualization you can set the viz flag (i.e., python leaderboard/scripts/run_evaluation.py user=$USER experiments=PlanTmedium3x eval=longest6 viz=1)

Explainability

The execution of the explainability agent contains two stages: (1) PlanT forwardpass (no execution of actions) to get attention weights. We filter the vehicles so that only the vehicles with the topk attention scores remain as input for the second step. (2) We execute either the expert or PlanT with the filtered input (the agent only sees topk vehicles instead of all).

Start a Carla server (see Data generation).
When the server is running, start the evaluation with:

python leaderboard/scripts/run_evaluation.py user=$USER experiments=PlanTExplainability experiments.exec_model=Expert experiments.topk=1

To obtain the attention visualization set experiments.topk=100000 and in addition add the flag save_explainability_viz=True. This saves a video per route in a viz_vid folder. The image resolution can be changed in carla_agent_files/explainability_agent.py.
Attention: saving the videos slows the evaluation down.

Perception PlanT

We release two PlanT agents suitable for the two CARLA Leaderboard tracks. For the SENSORS track we predict the route with our perception module. In the MAP track model we get the route information from the map. The code is taken from the TransFuser (PAMI 2022) repo and adapted for our usecase. The config is specified in the folder carla_agent_files/config. The config for the perception model is in training/Perception/config.py.

SENSORS track

Start a Carla server (see Data generation).

When the server is running, start the evaluation with:

python leaderboard/scripts/run_evaluation.py user=$USER experiments=PlanTSubmission track=SENSORS eval=longest6 save_path=SENSORSagent

Visualization can be activated with the viz flag, and the unblocking from the TransFuser repo can be activated with the experiments.unblock flag.

MAP track

Start a Carla server (see Data generation).

When the server is running, start the evaluation with:

python leaderboard/scripts/run_evaluation.py user=$USER experiments=PlanTSubmissionMap track=MAP eval=longest6 save_path=MAPagent

Visualization can be activated with the viz flag, and the unblocking from the TransFuser repo can be activated with the experiments.unblock flag.

Citation

If you use this code and data, please cite the following:

@inproceedings{Renz2022CORL,
    author       = {Katrin Renz and Kashyap Chitta and Otniel-Bogdan Mercea and A. Sophia Koepke and Zeynep Akata and Andreas Geiger},
    title        = {PlanT: Explainable Planning Transformers via Object-Level Representations},
    booktitle    = {Conference on Robotic Learning (CoRL)},
    year         = {2022}
}

Also, check out the code for other recent work on CARLA from our group:

• Chitta et al., TransFuser: Imitation with Transformer-Based Sensor Fusion for Autonomous Driving (PAMI 2022)
• Hanselmann et al., KING: Generating Safety-Critical Driving Scenarios for Robust Imitation via Kinematics Gradients (ECCV 2022)
• Chitta et al., NEAT: Neural Attention Fields for End-to-End Autonomous Driving (ICCV 2021)