README.md

metadata

license: apache-2.0
language:
  - en
tags:
  - Multi-agent
  - LLM
  - AI
pretty_name: O
size_categories:
  - 10B<n<100B

ORCH: Organizational Principles Enable Collective Intelligence in Embodied AI

Project Website | Video | Paper

Authors

Zhengran Ji, Jonathan Hyun, Boyuan Chen.

Duke University, General Robotics Lab

Overview

result.tar.gz contains experiment outputs for cooperative multi-agent wildfire simulation tasks. The data supports analysis of task performance, agent communication, model usage, and alternative team structures. It contains simulation results, rather than observational records of real wildfires.

The archive includes per-run CSV metrics, per-agent text conversation logs, and CSV tables comparing team structures. Its compressed size is approximately 30.57 GB (28.47 GiB); extraction requires additional disk space.

This README is based on sampled archive contents and inspection of the accompanying extracted result/ directory. It is not a complete archive integrity or completeness audit.

Directory Structure

The standard experiment layout is:

result/
├── <model>/
│   └── <method>/
│       └── <task>/
│           └── <seed>/
│               └── <timestamp>/
│                   ├── data.csv
│                   ├── Agent_1/
│                   │   └── chats.txt
│                   └── Agent_<n>/
│                       └── chats.txt
├── noop/
│   └── ...
└── team_structure_id_tables/
    └── *.csv

• Model: model-family label. Observed labels are deepseek, ernie, gemma, glm, gpt, llama, nemotron, and qwen. These labels alone do not identify exact model versions or inference settings.
• Method: experiment method or team-configuration variant.
• Task: scenario identifier.
• Seed: numeric experiment seed, such as 4792.
• Timestamp: run directory named in YYYY-MM-DD-HH-MM-SS format. The timezone is not specified.
• Agent directory: individual agent logs. Agent counts vary by scenario and configuration.
• noop/: separate baseline results organized under task directories; do not assume the standard model/method nesting applies.

A model/task/seed combination may contain multiple timestamped runs. Preserve the full path as the run identifier.

Methods and Team Configurations

Observed directory labels include:

Label	Description
CAMON, COELA, Embodied, HMAS_2	Comparison methods, using their stored names.
WILDFIRE	Main WILDFIRE experiment results.
WILDFIRE_default_teamconfig	Default team-configuration variant.
WILDFIRE_llm_generated_teamconfig	LLM-generated team-configuration variant.
WILDFIRE_llm_generated_teamconfig_no_critic	LLM-generated configuration variant without a critic.
WILDFIRE_llm_generated_only_horizontal	Horizontal-only configuration variant.
WILDFIRE_llm_generated_only_vertical	Vertical-only configuration variant.

Some model folders also contain directories ending in _old or _log, including COELA_old, WILDFIRE_old, and WILDFIRE_expert_preset_teamconfig_old. Treat these as separate variants until their relationship to the main runs is established. Not every model has every method or variant.

Tasks

The team-structure tables list the following 25 scenario identifiers. Their presence in a table does not establish complete run coverage for every model and method.

Cut_Trees_Sparse_small
Cut_Trees_Sparse_large
Cut_Trees_Lines_small
Cut_Trees_Lines_large
Scout_Fire_small
Scout_Fire_large
Transport_Firefighters_small
Transport_Firefighters_large
Rescue_Civilians_Known_Location_small
Rescue_Civilians_Known_Location_large
Suppress_Fire_Contain
Suppress_Fire_Extinguish
Rescue_Civilians_Search_and_Rescue
Suppress_Fire_Locate_and_Suppress
Suppress_Fire_Locate_Deploy_Suppress
Rescue_Civilians_Search_Rescue_Transport
Full_Game
Scout_Fire_Drone_Lost
Transport_Helicopter_Down
Rescue_Civilians_Surprise
Suppress_Fire_Extinguish_Second_Fire
Suppress_Fire_Contain_Water_Source
Suppress_Fire_Extinguish_Rapid_Growth
Scale_Level_Simple
Scale_Level_Complex

These cover tree cutting, scouting, firefighter transport, civilian rescue, fire suppression, combined objectives, disruptions, and different team scales.

Data Files

data.csv

Each sampled file is a comma-separated table with a header and time-indexed simulation metrics. The following columns were observed in a sampled run. Descriptions reflect field names; exact counter semantics, units, and reset behavior should be checked against the simulation implementation before deriving scores.

Column	Interpretation
timestep	Simulation step index.
exploration	Exploration metric.
trees_fire	Tree loss associated with fire.
trees_agents	Tree cutting associated with agents.
fire_extinguished	Fire-extinguishing metric.
correct_trees_cut	Trees cut toward the task objective.
civilians_rescued	Civilian rescue metric.
firefighters_transported	Firefighter transport metric.
civilians_scouted	Civilian discovery metric.
fire_scouted	Fire discovery metric.
water_scouted	Water-source discovery metric.
agents_destroyed	Agent loss metric.
civilians_destroyed	Civilian loss metric.
buildings_destroyed	Building loss metric.
cumulative_api_calls	Accumulated model API calls.
cumulative_input_tokens	Accumulated input tokens.
cumulative_output_tokens	Accumulated output tokens.
cumulative_cost	Accumulated recorded model cost; currency and pricing basis are not established here.
cumulative_idle_steps	Accumulated idle-step counter.
cumulative_replans	Accumulated replanning counter.
time	Recorded time value; verify units and timing semantics before use.

Inspect each file's header before combining runs: schemas may differ across methods or experiment versions. The final row is the last recorded state, not necessarily evidence of successful task completion.

Agent_<n>/chats.txt

Plain-text logs record agent prompts and generated responses. Sampled logs include observation summaries, team composition, task instructions, communication, and action selection. Sections may use labels such as Summarizing Observations, Choosing an Action, user, assistant, and END CHAT.

These files are not a standardized JSON conversation format. Log sections and formatting may vary by method. Generated statements are agent outputs and should be distinguished from simulator-recorded outcomes.

team_structure_id_tables/

File	Contents
team_structure_ids_<model>.csv	Task-level team-structure IDs across human presets and generated configuration variants. Some tables include a default column.
llm_generated_team_hierarchy_ids_7_open_source_models.csv	Task-level hierarchy IDs across seven model labels, with a distinct_team_hierarchies column.
expert_vs_llm_generated_vs_no_critic_team_ids_by_model.csv	Comparison of expert, generated, and no-critic team IDs by model.
invalid_horizontal_vertical_manager_configs.csv	Validation findings with model, task, variant, status, problem, and path columns.

Structure IDs are categorical labels, not performance scores. Do not assume their numerical values have a global meaning across tasks or tables without checking the ID-generation procedure. Preserve blank entries as missing values. Absolute paths in validation tables refer to the original working environment and may not resolve on another machine.

Getting Started

Run the following from the directory containing result.tar.gz.

List archive contents without extracting:

tar -tzf result.tar.gz

Extract into a chosen directory:

mkdir -p extracted
tar -xzf result.tar.gz -C extracted