PlanScope

This is the official repository of

PlanScope: Learning to Plan Within Decision Scope Does Matter

Ren Xin, Jie Cheng, Hongji Liu and Jun Ma

TL;NR

Based on PLUTO, we study the integrating method of long and short-term decision making, and the Time Dependent Normalization achieves the most significant improvement to 91.32% in the nuPlan Val4 CLS-NR score.

Video

Performance Comparison with SOTA methods

w/o post-processing | Model Name | Val14 CLS-NR Score | Val14 CLS-R Score | | ---- | ---- | ---- | | PLUTO | 89.04 | 80.01 | | STR2-CPKS-800M| 65.16 | - | | Diffusion Planner | 89.87 | 82.80 | | PlanScope (ours) | 91.32 | 80.96 |

Setup Environment

Setup dataset

Setup the nuPlan dataset following the official-doc

Setup conda environment

conda create -n planscope python=3.9
conda activate planscope

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

# setup planscope
cd ..
git clone https://github.com/Rex-sys-hk/PlanScope && cd planscope
sh ./script/setup_env.sh

Feature Cache

Preprocess the dataset to accelerate training. It is recommended to run a small sanity check to make sure everything is correctly setup.

 python run_training.py \
    py_func=cache +training=train_pluto \
    scenario_builder=nuplan_mini \
    cache.cache_path=/nuplan/exp/sanity_check \
    cache.cleanup_cache=true \
    scenario_filter=training_scenarios_tiny \
    worker=sequential

Then preprocess the whole nuPlan training set (this will take some time). You may need to change cache.cache_path to suit your condition

 export PYTHONPATH=$PYTHONPATH:$(pwd)

 python run_training.py \
    py_func=cache +training=train_pluto \
    scenario_builder=nuplan \
    cache.cache_path=/nuplan/exp/cache_pluto_1M \
    cache.cleanup_cache=true \
    scenario_filter=training_scenarios_1M \
    worker.threads_per_node=40

Training

sh train_scope.sh

• you can remove wandb related configurations if your prefer tensorboard.

Checkpoint

Copy your chekpoint path to sim_scope.sh or sim_pluto.sh and replace the value of CKPT_N to run the evaluation.

Run PlanScope-planner simulation

Run simulation for a random scenario in the nuPlan-mini split

sh ./sim_scope.sh

To Do

The code is under cleaning and will be released gradually.

• [ ] improve docs
• [x] training code
• [x] visualization
• [x] Scope-planner & checkpoint
• [x] feature builder & model
• [x] initial repo & paper

Citation

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

@misc{planscope,
      title={{PlanScope:} Learning to Plan Within Decision Scope Does Matter}, 
      author={Ren Xin and Jie Cheng and Jun Ma},
      year={2024},
      eprint={2411.00476},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2411.00476}, 
}

Thanks

• tuplan
• pluto

Others

• Please mind the common problem of nuPlan Dataset setup: https://github.com/motional/nuplan-devkit/issues/379
• Advised NATTEN Version: 0.14.6+torch1121cu116
• Please mind your Linux system version, Ubuntu 18.04.6 LTS is prefered. Debian may lead to some unexpected error in closed-loop simulation.
• When training on the 20% dataset, the random selection of data splits during training possibly cause fluctuations of about 2% CLS-NR score on Random14, the training on partial dataset should only be used as reference during development.
• This repo is updated on 5 March 2025, the previous version can be found by checkout branch archived_1.