GlucoSim

JAX-based diabetes simulation environments to test safe generalization across safe RL algorithms

Dependencies

Install the pinned minimums:

python -m pip install -r requirements.txt

Tested Python version: 3.10

This is simulator only. For those who want to test algorithms, please refer to https://github.com/safe-autonomy-lab/GlucoAlg

Why both gym_env and safety_gymnasium?

Both folders provide Gymnasium-style APIs with costs, but for different consumers:

• envs/gym_env: we need our local gym env because this environment should return reward and cost, so make(...) call that returns (obs, reward, cost, terminated, truncated, info).
• envs/safety_gymnasium: Safety-Gymnasium compatibility layer used by OmniSafe-style workflows. It is updated to work with Python >= 3.10, but OmniSafe itself is not compatible with Python >= 3.10 as of today.

When OmniSafe catches up to Python >= 3.10, use envs/diabetes_cmdp.py as the entry point for OmniSafe algorithms.

Package install

This repo is laid out as a Python package (envs/ as the top-level module). To use it as an installed package:

1. Add a minimal pyproject.toml (setuptools) at the repo root.
2. Run:

python -m pip install -e .

If you don't want to package it, you can also run from source by adding the repo root to PYTHONPATH.

Notes

• The safety_gymnasium package mirrors Gymnasium wrappers but preserves cost signals.
• The local gym_env registry is separate from Gymnasium's global registry; use from glucosim import gym_env as gym to get gym.make(...) behavior.

Quickstart (Gym-like usage)

The repo registers t1d-v0, t2d-v0, and t2d_no_pump-v0 on import. Use the local gym-like registry:

import glucosim  # registers t1d-v0 / t2d-v0 / t2d_no_pump-v0
from glucosim import gym_env as gym

# Minimum episode length is one day! so even if you set up 12 * 60 (half day) as an episode length, it would convert to one day.
env = gym.make(
    "t1d-v0",
    simulation_minutes=24 * 60,
    sample_time=5,
    patient_name="adolescent#001",
)

obs, info = env.reset()
for _ in range(10):
    action = env.action_space.sample()
    obs, reward, cost, terminated, truncated, info = env.step(action)
    if terminated or truncated:
        break

Key knobs:

• simulation_minutes: total episode length in minutes.
• sample_time: controller interval (minutes per step).
• patient_name: from envs/simglucose/params/vpatient_params.csv.

OmniSafe / CMDP usage

Use the CMDP wrapper in envs/diabetes_cmdp.py:

from glucosim.diabetes_cmdp import DiabetesEnvs

cmdp = DiabetesEnvs(
    env_id="t1d-v0",
    device="cpu",
    num_envs=1,
    simulation_minutes=24 * 60,
    sample_time=5,
    patient_name="adolescent#001",
)

obs, info = cmdp.reset()

This wrapper returns costs in the step API and is the intended entry point for OmniSafe algorithms.

Patient randomization / generalization

You can generate more patient variation without adding new CSVs by sampling patient_overrides at env creation time. These overrides are already supported by JaxSimEnv and create_patient_params.

Built-in scaling knobs you can randomize:

• carb_absorption_scale (scales kmax, kabs)
• insulin_sensitivity_scale (scales Vmx)
• autobalance_basal_scale / autobalance_hepatic_scale (shift basal steady-state)
• eat_rate_scale (meal intake rate)

Example:

import numpy as np
import glucosim
from glucosim import gym_env as gym

rng = np.random.default_rng(0)
overrides = {
    "carb_absorption_scale": rng.uniform(0.8, 1.2),
    "insulin_sensitivity_scale": rng.uniform(0.8, 1.2),
    "autobalance_basal_scale": rng.uniform(0.85, 1.15),
    "eat_rate_scale": rng.uniform(0.85, 1.15),
}

env = gym.make(
    "t1d-v0",
    patient_name="adult#001",
    patient_overrides=overrides,
)

Models and Datasets

We uploaded our models and datasets in the following links

If this repository was helpful to your research, please consider citing our work:

APA Style:

Kwon, M., Lamp, J., & Feng, L. (2026). Safety Generalization Under Distribution Shift in Safe Reinforcement Learning: A Diabetes Testbed. arXiv preprint arXiv:2601.21094.

BibTeX:

@misc{kwon2026safetygeneralizationdistributionshift,
      title={Safety Generalization Under Distribution Shift in Safe Reinforcement Learning: A Diabetes Testbed}, 
      author={Minjae Kwon and Josephine Lamp and Lu Feng},
      year={2026},
      eprint={2601.21094},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={[https://arxiv.org/abs/2601.21094](https://arxiv.org/abs/2601.21094)}, 
}