pltf: A Python repository from Yisten

This is the official repository of

Rethink Imitation-based Planner for Autonomous Driving, Jie Cheng,Yingbing chen,Xiaodong Mei,Bowen Yang,Bo Li and Ming Liu, arXiv 2023

Highlight

A good starting point for research on learning-based planner on the nuPlan dataset. This repo provides detailed instructions on data preprocess, training and benchmark.
A simple pure learning-based baseline model planTF, that achieves decent performance without any rule-based strategies or post-optimization.

Setup Environment

setup the nuPlan dataset following the offiical-doc
setup conda environment

conda create -n plantf python=3.9
conda activate plantf

# install nuplan-devkit
git clone https://github.com/motional/nuplan-devkit.git && cd nuplan-devkit
pip install -e .
pip install -r ./requirements.txt

# setup planTF
cd ..
git clone https://github.com/jchengai/planTF.git && cd planTF
sh ./script/setup_env.sh

Feature cache

Preprocess the dataset to accelerate training. The following command generates 1M frames of training data from the whole nuPlan training set. You may need:

change cache.cache_path to suit your condition
decrease/increase worker.threads_per_node depends on your RAM and CPU.

 export PYTHONPATH=$PYTHONPATH:$(pwd)

 python run_training.py \
    py_func=cache +training=train_planTF \
    scenario_builder=nuplan \
    cache.cache_path=/nuplan/exp/cache_plantf_1M \
    cache.cleanup_cache=true \
    scenario_filter=training_scenarios_1M \
    worker.threads_per_node=40

This process may take some time, be patient (20+hours in my setting).

Training

We modified the training scipt provided by nuplan-devkit a little bit for more flexible training. By default, the training script will use all visible GPUs for training. PlanTF is quite lightweight, which takes about 4~6G GPU memory under the batch size of 32 (each GPU).

CUDA_VISIBLE_DEVICES=0,1,2,3 python run_training.py \
  py_func=train +training=train_planTF \
  worker=single_machine_thread_pool worker.max_workers=32 \
  scenario_builder=nuplan cache.cache_path=/nuplan/exp/cache_plantf_1M cache.use_cache_without_dataset=true \
  data_loader.params.batch_size=32 data_loader.params.num_workers=32 \
  lr=1e-3 epochs=25 warmup_epochs=3 weight_decay=0.0001 \
  lightning.trainer.params.val_check_interval=0.5 \
  wandb.mode=online wandb.project=nuplan wandb.name=plantf

you can remove wandb related configurations if your prefer tensorboard.

Trained models

Place the trained models at planTF/checkpoints/

Model	Document	Download
PlanTF (state6+SDE)	-	OneDrive
RasterModel	Doc	OneDrive
UrbanDriver (openloop)	Doc	OneDrive

Evaluation

run a single scenario simulation (for sanity check): sh ./script/plantf_single_scenarios.sh
run Test14-random: sh ./script/plantf_benchmarks.sh test14-random
run Test14-hard: sh ./script/plantf_benchmarks.sh test14-hard
run Val14 (this may take a long time): sh ./script/plantf_benchmarks.sh val14

Results

Test14-random and Test14-hard benchmarks

Planners		Test14-random			Test14-hard
Type	Method	OLS↑	NR-CLS↑	R-CLS↑	OLS↑	NR-CLS↑	R-CLS↑	Time
Expert	LogReplay	100.0	94.03	75.86	100.0	85.96	68.80	-
Rule-based	IDM	34.15	70.39	72.42	20.07	56.16	62.26	32
Rule-based	PDM-Closed	46.32	90.05	91.64	26.43	65.07	75.18	140
Hybrid	GameFormer	79.35	80.80	79.31	75.27	66.59	68.83	443
Hybrid	PDM-Hybrid	82.21	90.20	91.56	73.81	65.95	75.79	152
Learning-based	PlanCNN	62.93	69.66	67.54	52.4	49.47	52.16	82
	UrbanDriver ^†	82.44	63.27	61.02	76.9	51.54	49.07	124
	GC-PGP	77.33	55.99	51.39	73.78	43.22	39.63	160
	PDM-Open	84.14	52.80	57.23	79.06	33.51	35.83	101
	PlanTF (Ours)	87.07	86.48	80.59	83.32	72.68	61.7	155

^† open-loop re-implementation

Val14 benchmark

Method	OLS	NR-CLS	R-CLS
Log-replay	100	94	80
IDM	38	77	76
GC-PGP	82	57	54
PlanCNN	64	73	72
PDM-Hybrid	84	93	92
PlanTF (Ours)	89.18	84.83	76.78

Acknowledgements

Many thanks to the open-source community, also checkout these works:

Citation

If you find this repo useful, please consider giving us a star 🌟 and citing our related paper.

@misc{cheng2023plantf,
      title={Rethinking Imitation-based Planner for Autonomous Driving},
      author={Jie Cheng and Yingbing Chen and Xiaodong Mei and Bowen Yang and Bo Li and Ming Liu},
      year={2023},
      eprint={2309.10443},
      archivePrefix={arXiv},
      primaryClass={cs.RO}
}

Yisten/pltf