CoCo — The Constitutional Controller

Resin reasoning · ProMis mission design · CoFi tracking · CoCo control

The Constitutional Controller: Doubt-Calibrated Steering of Compliant Agents
Simon Kohaut, Felix Divo, Navid Hamid, Benedict Flade, Julian Eggert, Devendra Singh Dhami, Kristian Kersting
Paper · Project Page · Citation

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.

The rules behind the landscape are written in Resin; see its language reference for their syntax and semantics. Paths can then be evaluated or steered to remain constitutionally compliant even under uncertainty. Since CoCo steers agents by a known constitution, other agents can rely on it: CoFi tracks the drones above from coarse, noisy observations by trusting them to follow the very same rules.

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. Learn a doubt density

import torch
from coco import DoubtDensity

# Control errors in meters, which grow with the velocity the agent flies at
velocity = torch.rand(2_000, 1)  # m/s
errors = torch.randn(2_000, 2) * (0.05 + 0.2 * velocity)

# The doubt space: the features that condition the agent's inaccuracies
doubt_space = {"velocity": {"type": "continuous", "values": velocity}}

density = DoubtDensity(doubt_space, number_of_states=2, number_of_hidden_features=32, number_of_layers=4)
losses = density.fit(errors, doubt_space, number_of_epochs=20, batch_size=64)
density.save("doubt_density.pkl")

2. Compute where the agent is allowed to be

import time
import numpy as np
from promis import ProMis, StaRMap
from promis.geo import CartesianMap, CartesianPolygon, CartesianRasterBand, PolarLocation

# The constitution: stay off the obstacle
constitution = """
over(obstacle) <- source("/star_map/over/obstacle", Probability).
landscape if not over(obstacle).
landscape -> target("/landscape").
"""

# A 2 m x 1 m obstacle in a 10 m x 10 m area
obstacle = CartesianPolygon.from_numpy(np.array([[-1, 1, 1, -1], [-0.5, -0.5, 0.5, 0.5]]), location_type="obstacle")
area = CartesianMap(PolarLocation(0, 0), {"obstacle": [obstacle]})
points = CartesianRasterBand(PolarLocation(0, 0), (50, 50), 10, 10)
star_map = StaRMap(area)
star_map.initialize(points, number_of_random_maps=1, logic=constitution)
promis = ProMis(star_map, constitution, 50 * 50)
promis.initialize(points)
time.sleep(0.05)  # Resin's writers deliver values asynchronously
landscape = promis.update()

3. Apply doubt to the landscape

from coco import ConstitutionalController, DoubtDensity

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

# Where the agent may go when flying at 1 m/s, given where its controller will actually take it
query = {"velocity": {"type": "continuous", "values": torch.tensor([[1.0]])}}
doubtful_landscape = controller.apply_doubt(landscape, density, query, number_of_samples=100, scaler=1.0)

The doubtful landscape spreads each obstacle by where the controller may actually take the agent; scaler converts the doubt's units to the landscape's.

4. Evaluate path compliance

# A path passing 30 cm above the obstacle, as (T, 2) positions in meters
path = np.stack([np.linspace(-3, 3, 13), np.full(13, 0.8)]).T
compliance = controller.compliance(path, landscape, density, query, number_of_samples=500)
# compliance: (T,) probability of complying at each waypoint

Citation

If you use CoCo, please cite the paper below. CoCo builds on ProMis.

@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}
}