Code for Arza et al., Performance-guided Task-specific Optimization for Multirotor Design. This repository is a more efficient rewrite in Newton of the methodology. Getting the result below (evaluating 800 designs + baselines) took 27 h on one RTX 5090 GPU.
You need Linux and an NVIDIA GPU with driver ≥ 580 (the environment installs PyTorch built for CUDA 13.0). Install Miniconda if you don't have conda:
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh -b -p ~/miniconda3
~/miniconda3/bin/conda init bash
source ~/.bashrcThen,
git clone https://github.com/ntnu-arl/task-specific-multirotor-optimization.git
cd task-specific-multirotor-optimization
conda env create -f environment.yml
conda activate multirotor-optThe first run compiles the Warp kernels, which takes a minute.
Run all commands from the repository folder with the environment active (run conda activate multirotor-opt to activate the environment).
python evolve.py --run-dir results/myrun # the paper's budget: 400 BO + 400 CMA-ES.
python evolve.py --run-dir results/test --bo-evals 40 --cma-evals 40 # a short test run that should finish in a few hours.This optimizes for task B of the paper. For task A, set TASK=A on every command of the run (the default budget and seed are the same):
TASK=A python evolve.py --run-dir results/task_AThis should produce results/<run>/ which contains:
| File | Contents |
|---|---|
progress.csv, evaluations.jsonl |
one line per evaluated design |
progress.pdf |
fitness over evaluations and time |
final_design.json |
the final design |
final_comparison.mp4, final_report.pdf, final_metrics.json |
video and report of the final design against the baselines |
A run interrupted after its BO goes on in its own folder (with the same TASK):
python evolve.py --run-dir results/myrun --resumepython render_evaluation.py --design results/myrun/final_design.json --out results/myrun/render
python render_evaluation.py --design planar --out results/renders/planar # or franchi, rajappa, shenpython compare.py --out results/compare --seeds 8 --seed 1 \
--designs planar franchi rajappa shen paper results/myrun/final_design.jsonTrains every design 8 times from scratch (sequential halving each time) and evaluates each picked policy on 40960
episodes. paper is the optimized task B design reported on the paper obtained in the legacy codebase.
Out of memory error.
With the current setup, rl.py trains 80 controllers at once (TRAIN_CHUNK), which needs about 14 GB of GPU memory. On a smaller GPU, set
it lower. It is slower, and the results could change with different parameters.
TRAIN_CHUNK=20 python evolve.py --run-dir results/myrun@misc{arza2025performance,
title = {Performance-guided Task-specific Optimization for Multirotor Design},
author = {Etor Arza and Welf Rehberg and Philipp Weiss and Mihir Kulkarni and Kostas Alexis},
year = {2025},
eprint = {2510.04724},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2510.04724}
}