Superhuman Safe and Agile Racing through Multi-Agent Reinforcement Learning

Training and headless evaluation for multi-agent drone racing, with SA, IPPO, SP, LP and LP-MLP. Requires access to Agilicious internal; follow the access-request instructions in the Agilicious repository.

Installation

Use Linux, Python 3.9, CMake, a C++17 compiler, OpenMP and Eigen 3.4 headers. Run from this folder, with Agilicious in a separate checkout:

git clone git@github.com:uzh-rpg/agilicious_internal.git ../agilicious_internal
git -C ../agilicious_internal checkout 02696bc1d5a97d8aed5a3e9fb7bdaf381c6b6041
python3.9 -m venv .venv
source .venv/bin/activate
bash setup.sh ../agilicious_internal

Setup applies the simulator patch and builds locally. Dependencies go in .deps/; compiled files go in build/.

Training

bash run.sh train lp --output runs/lp

Choose sa, ippo, sp, lp or lp_mlp. For a short training check, add --num-envs 2 --threads 1 --steps 4 --epochs 1 --timesteps 16 --device cpu.

Evaluation

bash run.sh evaluate --controllers lp,lp_mlp,sa,ippo --output runs/mixed
bash run.sh benchmark self --output runs/self --workers 4 --plot
bash run.sh benchmark pool --output runs/pool --workers 4

Self evaluation repeats each checkpoint across 1–8 agents and 64 starts. Pool evaluation samples 1,000 four-agent lineups with equal method probability, then evaluates 64 starts each. Use --starts 1 for fewer starts, --lineups 1 for a single pool lineup, or --agents 4 to restrict self races.

--plot overlays each lineup's races with a different color per agent start slot. To plot existing results, run bash run.sh plot runs/self. Race outputs are in races/, logs in logs/, and summaries and plots at the top level.

Races run at 50 Hz for at most 20 s; completion requires 22 gate passages. summary.csv reports completion, gate fraction, rank, collisions and the mean of each participant's best completed simulator lap. Raw results retain corrected flying laps and trajectories. Downwash is stochastic and is not fully controlled by the evaluation seed. Rankings use deterministic ties, finishing order and frozen crash progress.

Checkpoints

Each method has four evaluated checkpoints: checkpoints/SA/1.pth–4.pth, with the same layout for SP, LP and LP-MLP. IPPO uses A0.pth–A3.pth for its four agents. Extra fixed training opponents comprise 3 SA policies in SA/training/ and 12 IPPO policies in IPPO/training/, named 1_a0.pth through 3_a3.pth (checkpoint number and agent). The 35 weight files include all evaluated policies and the opponents needed for LP/LP-MLP training.

License

GPLv3. See NOTICE.md for third-party licenses.

Citation

If you use this code, please cite:

@misc{geles2026superhumanmultiagentracing,
  title={Superhuman Safe and Agile Racing through Multi-Agent Reinforcement Learning},
  author={Ismail Geles and Leonard Bauersfeld and Markus Wulfmeier and Davide Scaramuzza},
  year={2026},
  eprint={2605.22748},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2605.22748},
}