PlanT 2.0 is a lightweight, object-centric planning transformer for CARLA that enables controlled analysis of failure modes while achieving strong closed-loop performance on CARLA Leaderboard 2.0 benchmarks.

• Object-centric input representation that can be perturbed for systematic analysis.
• Upgrades for CARLA Leaderboard 2.0 scenarios.
• State-of-the-art results on Longest6 v2, Bench2Drive, and CARLA validation routes.
• Open-sourced code, models, and dataset for reproducibility.

• Installation
• Quick start inference
• Evaluation at scale
• Training
• Dataset
• Project structure
• Citation
• Acknowledgments

# 1. Clone this repository
git clone https://github.com/autonomousvision/plant2.git
cd plant2

# 2. Setup CARLA (skip if you already have a compatible install)
chmod +x setup_carla.sh
./setup_carla.sh

# 3. Setup environment
conda env create -f environment.yml
conda activate plant2

You can test the performance of the pretrained models using the CARLA leaderboard evaluator.

Set the environment variables:

export CARLA_ROOT=/path/to/CARLA
export WORK_DIR=/path/to/plant2
export SCENARIO_RUNNER_ROOT=$WORK_DIR/scenario_runner_autopilot
export LEADERBOARD_ROOT=$WORK_DIR/leaderboard_autopilot
export PYTHONPATH=$CARLA_ROOT/PythonAPI/carla:$LEADERBOARD_ROOT:$SCENARIO_RUNNER_ROOT:$WORK_DIR/PlanT:$WORK_DIR/carla_garage

# Required by PlanT/PlanT_agent.py
export PLANT_CHECKPOINT=/path/to/checkpoint.ckpt
export PLANT_VIZ=/path/to/viz_outputs  # set to empty string to disable

Then start CARLA and run the evaluator:

# Terminal 1
$CARLA_ROOT/CarlaUE4.sh

# Terminal 2
python leaderboard_autopilot/leaderboard/leaderboard_evaluator_local.py \
  --routes=/path/to/route_file.xml \
  --track=[SENSORS if B2D else MAP] \
  --agent=PlanT/PlanT_agent.py \
  --checkpoint=results.json \
  --timeout=300

For SLURM-based evaluation of a benchmark, use PlanT/plant_evaluate.py. It generates per-route job scripts and handles retries.

Before running:

• Update the cfg block inside PlanT/plant_evaluate.py with your local paths and CARLA settings.
• Adjust PlanT/eval_num_jobs.txt to control the maximum parallel jobs.

Example:

python PlanT/plant_evaluate.py \
  --checkpoint /path/to/epoch=029_final_1.ckpt \
  --routes /path/to/longest6_split/ \
  --out_root results/longest6 \
  --seeds 1 2 3

Training uses Hydra configs plus a few environment variables.

1. Create a user config:

• Copy PlanT/config/user/simon.yaml to PlanT/config/user/<your_name>.yaml.
• Update working_dir to your local checkout.

2. Update PlanT/config/config.yaml:

• Replace TODO_your_username with your user file name.

3. Set required env vars and start training:

export SEED=1
export DS=/path/to/plant_dataset   # folder that contains a `data/` subdir
export DS_LOCAL=/path/to/fast_cache # optional, for diskcache
export CHECKPOINT_ADDON=run1        # optional

python PlanT/lit_train.py

For SLURM, PlanT/train_plant.sh shows an example job script. It is cluster-specific and expects a local dataset cache, so adapt the paths and SLURM settings to your environment.

The dataset used in the paper is available on Hugging Face: https://huggingface.co/datasets/SimonGer/PlanT2_Dataset.

To collect your own dataset:

• Use the modified carla_garage/autopilot.py and carla_garage/data_agent.py.
• Run the SLURM helpers in collect_dataset_slurm.py or 0_run_collect_dataset_slurm.sh (both are cluster-specific and require path/partition updates).
• Control parallel job limits via max_num_jobs.txt.
• The default collection settings do not require GPUs and collect object-level data only (no RGB).

• PlanT/ core model, agent, configs, and training code.
• carla_garage/ data collection agents and utilities.
• leaderboard_autopilot/ and scenario_runner_autopilot/ CARLA Leaderboard 2.0 forks used for evaluation.
• Bench2Drive/ assets and scripts for Bench2Drive evaluation.

@misc{gerstenecker2025plant20exposingbiases,
      title={PlanT 2.0: Exposing Biases and Structural Flaws in Closed-Loop Driving},
      author={Simon Gerstenecker and Andreas Geiger and Katrin Renz},
      year={2025},
      eprint={2511.07292},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2511.07292},
}

This repository builds upon the work of carla_garage and PlanT. We thank the authors for open-sourcing their work.