Cooperative Navigation Suite: A Testbed for Emergent Communication in Asymmetric Multi-Agent Systems
conav_suite is a research platform for studying emergent communication protocols in cooperative multi-agent systems. Built as an extension of the Multi-Agent Particle Environment (MPE) from the Farama Foundation, conav_suite provides a flexible testbed for developing and evaluating communication strategies between agents with asymmetric information.
- Asymmetric Information : Aerial agents with global vision and ground agents with local perception
- Diverse Scenarios : Eight unique problem configurations with varying obstacle arrangements
- Dynamic Environments : Configurable obstacle placements and agent/goal positions
- Communication Challenge : No predefined communication protocol—agents must develop their own language
- PettingZoo Compatible : Follows the PettingZoo API for easy integration
In conav_suite, agents must navigate to goal positions while avoiding obstacles. The core challenge emerges from the asymmetric information structure:
- Aerial Agent : Has global visibility of all obstacles (both large and small)
- Ground Agent : Can only perceive small obstacles in its vicinity
Large obstacles are constrained to specific regions defined by each problem instance, while small obstacles can appear anywhere in the environment. Start and goal positions are placed outside the large obstacle regions.
The aerial agent must develop effective communication strategies to guide the ground agent safely to its goal, without a pre-defined communication protocol.
conav_suite draws inspiration from the Lewis Signaling Game , a game-theoretic framework for studying the emergence of communication. The environment serves as a testbed for several open research questions:
- How do agents develop communication protocols without pre-defined structures?
- What makes communication efficient in asymmetric information settings?
- How do different environmental configurations affect emergent communication?
- Can agents develop context-dependent language and high-level directives?
These questions have applications in robotics, autonomous systems coordination, human-AI interaction, and fundamental understanding of communication emergence.
pip install conav-suitegit clone https://github.com/ethanmclark1/conav_suite.git
cd conav_suite
pip install -e .Required packages are automatically installed with the above methods, but can also be installed manually:
conda env create -f environment.ymlimport conav_suite
# Create environment
env = conav_suite.env()
# Reset environment with a specific problem instance
env.reset(options={'problem_instance': 'bisect'})
# Get initial observation
observation, reward, termination, truncation, info = env.last()
# Take a step in the environment
action = env.action_space(env.agent_selection).sample() # Replace with your agent's policy
env.step(action)
# Close environment when done
env.close()conav_suite offers eight distinct problem configurations that define constraint regions for large obstacle placement:
| Problem Instance | Visualization |
|---|---|
bisect |
|
circle |
|
corners |
|
cross |
|
staggered |
|
quarters |
|
stellaris |
|
scatter |
Note: The red regions in these visualizations represent areas where large obstacles can appear. Agents, goals, and small obstacles can be placed in any non-red area. This creates distinct navigational challenges across different problem instances, as the aerial agent must effectively communicate the locations of large obstacles (which only it can see) to the ground agent.
conav_suite follows the PettingZoo API. Here are the key components:
env = conav_suite.env()env = conav_suite.env(
num_agents=1, # Number of ground agents
num_large_obstacles=4, # Number of large obstacles
large_obstacle_radius=0.05, # Size of large obstacles
num_small_obstacles=10, # Number of small obstacles
small_obstacle_radius=0.02, # Size of small obstacles
render_mode=None, # None, "human", or "rgb_array"
max_cycles=100 # Maximum steps per episode
)- Ground Agent : Position (2), goal position (2), other agent positions (2 × num_agents-1), small obstacle positions (2 × num_small_obstacles)
- Aerial Agent : Full state observation of all entities
For more details, refer to the PettingZoo API documentation.
We welcome contributions to conav_suite! You can contribute in several ways:
- Bug Reports : If you discover a bug, please submit a report via the issues tab with a detailed description and, if possible, a code sample.
- Feature Requests : Have ideas to improve conav_suite? Post your suggestions in the issues tab with a detailed explanation.
- Pull Requests : Made enhancements? Submit a pull request! We appreciate all help to make conav_suite better.
If you encounter issues or have questions, please:
- Create an issue in the GitHub repository
- Contact me directly at eclark715@gmail.com
conav_suite is open-source software licensed under the MIT license.
Our environment builds upon these foundational works:
@article{lowe2017multi,
title={Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments},
author={Lowe, Ryan and Wu, Yi and Tamar, Aviv and Harb, Jean and Abbeel, Pieter and Mordatch, Igor},
journal={Neural Information Processing Systems (NIPS)},
year={2017}
}
@article{mordatch2017emergence,
title={Emergence of Grounded Compositional Language in Multi-Agent Populations},
author={Mordatch, Igor and Abbeel, Pieter},
journal={arXiv preprint arXiv:1703.04908},
year={2017}
}