CoCo — The Constitutional Controller

Doubt-Calibrated Steering of Compliant Agents

---

CoCo couples low-level motion control with probabilistic logic by learning a self-doubt density, a Conditional Normalizing Flow that models an agent's control inaccuracies as a function of its operating conditions (e.g. velocity, heading). This doubt distribution is folded into a ProMis compliance landscape at query time, producing a doubt-calibrated landscape that accounts for where the agent will actually end up, not just where it intends to go. Paths can then be evaluated or steered to remain constitutionally compliant even under uncertainty.

Installation

pip install python-coco

To install from source:

git clone https://github.com/simon-kohaut/CoCo
cd CoCo
pip install -e .

Quick Start

1. Train a doubt density

import numpy as np
from coco import DoubtDensity

# Define the doubt space: features that condition the agent's inaccuracies
doubt_space = {
    "velocity": {
        "type": "continuous",
        "values": np.array([...]),   # shape (N,) — one value per training sample
    }
}

# Observed control errors (shape N x 2)
samples = np.array([...])

density = DoubtDensity(
    doubt_space=doubt_space,
    number_of_states=2,
    number_of_hidden_features=32,
    number_of_layers=4,
)
losses = density.fit(samples, doubt_space, number_of_epochs=100, batch_size=64)

density.save("doubt_density.pkl")

2. Apply doubt to a compliance landscape

from coco import ConstitutionalController, DoubtDensity
from promis import ProMis, StaRMap

density = DoubtDensity.load("doubt_density.pkl")

# doubt_space at inference time (query conditions)
query_doubt_space = {
    "velocity": {"type": "continuous", "values": np.array([target_speed])}
}

# landscape is a ProMis CartesianCollection with compliance values
controller = ConstitutionalController()
doubtful_landscape = controller.apply_doubt(
    landscape=landscape,
    doubt_density=density,
    doubt_space=query_doubt_space,
    number_of_samples=500,
)

3. Evaluate path compliance

# path: np.ndarray of shape (T, 2)
compliance_scores = controller.compliance(
    path=path,
    landscape=landscape,
    doubt_density=density,
    doubt_space=query_doubt_space,
    number_of_samples=500,
)
# compliance_scores: (T,) array in [0, 1]

Citation

@inproceedings{kohaut2026coco,
  title     = {The Constitutional Controller: Doubt-Calibrated Steering of Compliant Agents},
  author    = {Kohaut, Simon and Divo, Felix and Hamid, Navid and Flade, Benedict
               and Eggert, Julian and Dhami, Devendra Singh and Kersting, Kristian},
  booktitle = {Proceedings of the IEEE/RSJ International Conference on Intelligent
               Robots and Systems (IROS)},
  year      = {2026}
}