TuneControl

A collection of black-box controller-tuning problems with a common Python interface. Choose a problem and objective, supply controller gains, and receive a scalar cost plus the simulated trajectory.

The collection currently includes CartPole and cascaded tanks, each with several objectives and deterministic or noisy configurations. You can use the problems with your own optimizer or add a new problem family.

Install

TuneControl currently supports Python 3.12. Install from GitHub:

python -m pip install "tunecontrol @ git+https://github.com/Data-Science-in-Mechanical-Engineering/tunecontrol.git"

For the Bayesian optimization notebook, include BoTorch:

python -m pip install "tunecontrol[botorch] @ git+https://github.com/Data-Science-in-Mechanical-Engineering/tunecontrol.git"

For development from a local checkout, run python -m pip install -e '.[dev,botorch]'.

Evaluate a controller

from tunecontrol import CartPole, CartPoleConfig

problem = CartPole(CartPoleConfig(dim=2, objective="mae"))
theta = problem.bounds.mean(dim=0)
value, info = problem.evaluate(theta)

print(f"Cost: {value.item():.4f}")
print("Controller:", theta.tolist())
print("State columns:", info["trajectory"]["state_names"])

Expected output:

Cost: 8.2734
Controller: [-40.0, -6.25]
State columns: ('cart_position', 'cart_velocity', 'pole_angle', 'pole_angular_velocity')

The bounds describe the controller search space; their midpoint is an example controller. info["trajectory"] contains timestamps, state and input arrays, and column names and units. See evaluate and plot to visualize both built-in problems.

Problems

Both families provide deterministic and noisy configurations. Standard configurations enumerate combinations of dimensions, objectives, and noise; custom settings are also supported. The linked problem descriptions define the models, objectives, bounds, and sampling conventions.

import tunecontrol as tc

print(tc.list_problems())  # ['cartpole', 'cascaded_tank']
configs = tc.available_configs("cartpole")
problem = tc.make("cartpole", config={"dim": 2, "objective": "mae"})

Direct Python construction and registry construction use the same configurations. See discover and configure for configuration serialization and reconstruction.

Darker colors indicate lower cost. White regions mark undefined evaluations.

Noise and evaluation behavior

Configurations are deterministic by default (noise=None). To add process noise and CartPole initial-state perturbations, supply a noise configuration:

from tunecontrol import CartPole, CartPoleConfig, CartPoleNoise

problem = CartPole(CartPoleConfig(dim=2, noise=CartPoleNoise()))
problem.setup(run_seed=42)
value, info = problem.evaluate(problem.bounds.mean(dim=0))

Each problem owns its random stream. Successive noisy evaluations advance that stream; calling setup(run_seed=42) again replays the sequence. Record the family, full configuration, package version, and seed when saving an experiment. Seeded replay assumes the same software and runtime environment.

evaluate accepts a finite floating-point tensor of shape (problem.dim,). Gains must lie within the declared bounds, including the endpoints. Out-of-bounds gains raise ValueError before simulation. Other invalid inputs also raise errors. Undefined or nonfinite objectives return (NaN, info) with diagnostics preserved. Check torch.isnan(value) before using a cost in an optimizer. Penalty handling belongs to the optimizer or example, rather than the problem itself.

Examples

Start with the example guide. Notebooks include saved outputs and plots so you can read them directly on GitHub.

Contributing

New controller-tuning problems are welcome. Start with the custom-task notebook and the family API and contribution guide. Document the model, controller parameters, objective definitions, noise, and sampling, and include tests and an example trajectory with your contribution.

Citation

TuneControl was introduced in A Decade of Bayesian Optimization for Controller Tuning and Robot Learning: Tutorial, Review, and Future Prospects. Please cite it if TuneControl supports your research.

@misc{stenger2026decade,
    title = {A Decade of {Bayesian} Optimization for Controller Tuning and Robot Learning: Tutorial, Review, and Future Prospects},
    author = {Stenger, David and Brunzema, Paul and Menn, Johanna and von Rohr, Alexander and Schoellig, Angela P. and Trimpe, Sebastian},
    year = {2026},
    eprint = {2609.09403},
    archivePrefix = {arXiv},
    primaryClass = {cs.RO},
    doi = {10.48550/arXiv.2609.09403},
    url = {https://arxiv.org/abs/2609.09403}
}

License

Released under the MIT License.