Collaborative Multi-Agent Block-Pushing Environment for Collective Planning with LLM Agents

Heads up! CUBE is currently undergoing cleanup to make the codebase more polished and developer-friendly. The environment is functional, but some components are still being streamlined. Feel free to explore and experiment, keeping in mind that things are still evolving.

CUBE is a lightweight, scalable, and interpretable environment for studying cooperative intelligence in LLM agents, RL agents, and hybrid neurosymbolic systems at scale.

Existing multi-agent benchmarks often involve only a few agents or require many agents without meaningful cognitive demands. CUBE fills this gap by introducing tasks where difficulty grows with the number of agents, requiring individual reasoning ability and genuine multi-agent cooperation.

Agents operate in a 2D grid world and must push weighted blocks into a goal region while coordinating under embodied constraints such as:

• force requirements
• block-chain
• agent-chain
• dynamic environmental uncertainty

CUBE uses a dual-layer design that exposes and tests cooperative cognition:

• A primitive layer governing grid actions and physical dynamics
• A symbolic layer offering higher-level actions such as move_to_block, rendezvous, push_block, and wait_agents, making it natural for LLM-based planning and communication

A single parameter n controls grid size, team size, block distribution, and overall complexity, creating a transparent and reproducible difficulty curriculum for studying cooperation from small teams to hundreds of agents.

CUBE is also the environment used by DR. WELL, a decentralized neurosymbolic framework that tackles cooperation through joint negotiation, individual planning, and a shared world model for iterative improvement.

CUBE: Collaborative Multi-Agent Block-Pushing Environment for Collective Planning with LLM Agents

https://happyeureka.github.io/cube/

DR. WELL: Dynamic Reasoning and Learning with Symbolic World Model for Embodied LLM-Based Multi-Agent Collaboration

https://narjesno.github.io/DR.WELL/

• Cooperative intelligence at scale with hundreds of LLM agents
• Symbolic and primitive action layers supporting neurosymbolic reasoning
• Difficulty curriculum controlled by a single parameter
• Rich symbolic feedback for customizable supervision and analysis

If you use CUBE in your research, please cite:

@inproceedings{yangcube,
  title     = {CUBE: Collaborative Multi-Agent Block-Pushing Environment for Collective Planning with LLM Agents},
  author    = {Yang, Hanqing and Nourzad, Narjes and Chen, Shiyu and Joe-Wong, Carlee},
  booktitle = {Workshop on Scaling Environments for Agents},
  year      = {2025}
}

If you are exploring embodied LLM-based cooperation, you may also find DR. WELL relevant:

@inproceedings{nourzad2025drwell,
  title     = {DR. WELL: Dynamic Reasoning and Learning with Symbolic World Model for Embodied LLM-Based Multi-Agent Collaboration},
  author    = {Nourzad, Narjes and Yang, Hanqing and Chen, Shiyu and Joe-Wong, Carlee},
  booktitle = {Workshop on Bridging Language, Agent, and World Models for Reasoning and Planning},
  year      = {2025}
}