CE-CM: Partner Capability Estimation for Task-Agnostic Adaptation in Ad-Hoc Teamwork

This repository is an archive of the code used for the submission of the paper entitled Partner Capability Estimation for Task-Agnostic Adaptation in Ad-Hoc Teamwork, by Peter Tisnikar, Maja Swieczkowska, Benteng Ma, Gerard Canal, and Matteo Leonetti, which is available here. It contains two independent experiment codebases:

• TidyUP/: symbolic capability learning experiments in a PDDL cleanup domain.
• Overcooked/: capability inference and planning experiments in macro-action Overcooked.

The two folders are intentionally kept as separate runnable projects that can be installed and ran separately. Each has its own pyproject.toml, uv.lock, tests, and README. There is no shared root Python environment.

Repository Layout

CE-CM/
  TidyUP/
    README.md
    pyproject.toml
    uv.lock
    cecm/
    configs/
    domains/
    otpl/
    tests/
  Overcooked/
    README.md
    pyproject.toml
    uv.lock
    method/
    planners/
    precompute/
    experiment_configs/
    generated_levels/
    tests/

TidyUP

TidyUP/ contains the symbolic planning experiments.

Set up and test:

cd TidyUP
uv python install 3.11
uv sync
uv run python main.py --help
uv run pytest -q

Run a small experiment:

uv run python main.py --human_config only_wash_dense.json --n_problems 5

Planner compatibility: the bundled popf binary is expected to work only on compatible Linux systems. On other systems, install or build POPF separately and pass it with --planner_path /path/to/popf, or set POPF_PATH.

See TidyUP/README.md for the full usage notes.

Overcooked

Overcooked/ contains the macro-action Overcooked experiments. Its main external dependency is gym-macro-overcooked, installed by UV from the paper-release tag configured in Overcooked/pyproject.toml.

Set up and test:

cd Overcooked
uv python install 3.11
uv sync
uv run pytest -q

Run a small cached agent experiment after downloading caches:

./download_precomputed_data.sh
uv run python agent_experiment.py --type 1 --cached --runs 1

See Overcooked/README.md for the full experiment, cache, and human-data instructions.

Downloadable Data

Large generated artifacts are intentionally not stored in this repository.

For Overcooked, precomputed Q-table policies, rollout caches, diverse-rollout caches, and transition trees are downloadable:

cd Overcooked
./download_precomputed_data.sh

The human trajectory dataset is also downloadable:

cd Overcooked
./download_trajectories.sh

The downloader scripts fetch the public artifacts documented in Overcooked/README.md.

Citation

If you use any of the above in your work, please cite this work as:

@article{tisnikar2026partner,
title={Partner Capability Estimation for Task-Agnostic Adaptation in Ad-Hoc Teamwork},
author={Peter Tisnikar and Maja Swieczkowska and Benteng Ma and Gerard Canal and Matteo Leonetti},
year={2026},
eprint={2607.27177},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2607.27177},
}