A natural language Gym environment for goal-conditioned reinforcement learning
LittleZoo is a Gym environment built on top of the Playground environment. It enables agents to interact through natural language, where actions are given as text commands, and observations are returned as textual scene descriptions. This makes it an ideal testbed for LLM agents.
First, create a Conda environment:
conda create -n littlezoo python=3.10
conda activate littlezooThen, install the dependencies:
git clone https://github.com/LorisGaven/LittleZoo.git
cd LittleZoo
pip install -e .The environment provides observations in a structured textual format, along with additional information:
-
Observation (str):
You see: <obj_1>, <obj_2>, <obj_3>, <obj_4> You are standing on: <obj_i> | "nothing" Inventory (n/2): <obj_i>, <obj_j> | "empty" -
Info (dict):
{ 'goal': <goal>, 'possible_actions': [ 'Grasp', 'Release <obj_1>', 'Release <obj_2>', 'Release both', 'Go to <obj_1>', 'Go to <obj_2>', ..., 'Go to <obj_n>' ] }
Agents interact by passing a string to the step() function. Valid actions include:
"Go to <obj_i>"→ Move to the object<obj_i>."Grasp"→ Pick up an object if inventory space allows."Release <obj_i>"→ Drop a specific object."Release both"→ Drop all held objects.
LittleZoo offers four categories of hierarchical goals, increasing in complexity:
- Grasp {obj} → Navigate and grasp an object.
- Grow {plant} → Provide water to grow a plant.
- Grow {herbivore} → Grow a plant, then feed it to a herbivore.
- Grow {carnivore} → Grow a herbivore, then feed it to a carnivore.
-
Reward Structure:
- Achieving a goal grants +1 reward and ends the episode.
- Otherwise, the reward is 0.
-
Step Limits:
- Grasp: 3 steps
- Grow Plant: 6 steps
- Grow Herbivore: 10 steps
- Grow Carnivore: 15 steps
- Impossible goals (e.g., "Grow table"): 10 steps
Door, Chair, Desk, Lamp, Table, Cupboard, Sofa, Bookshelf, Bed
Carrot, Potato, Berry, Lettuce, Tomato, Cucumber, Spinach, Broccoli, Onion
Cow, Elephant, Rabbit, Deer, Sheep, Giraffe, Goat, Horse, Bison
Lion, Tiger, Bobcat, Panther, Coyote, Wolf, Leopard, Hyena, Jackal
Water
If you use LittleZoo in your research, please cite:
@article{gaven2025magellan,
title={MAGELLAN: Metacognitive predictions of learning progress guide autotelic LLM agents in large goal spaces},
author={Gaven, Loris and Carta, Thomas and Romac, Cl{\'e}ment and Colas, C{\'e}dric and Lamprier, Sylvain and Sigaud, Olivier and Oudeyer, Pierre-Yves},
journal={arXiv preprint arXiv:2502.07709},
year={2025}
}If you’d like to contribute to LittleZoo, feel free to open an issue or submit a pull request on GitHub. 🚀