MARGoS: A Configuration-Driven Platform for Reproducible MARL Experiments in ARGoS

Tom Philipp Backert; Mohammad D. Soorati; Javad Ghofrani

cs.SEOct 9, 2026

MARGoS: A Configuration-Driven Platform for Reproducible MARL Experiments in ARGoS

Authors: Tom Philipp Backert, Mohammad D. Soorati, Javad Ghofrani

Organizations: University of Lübeck, Lübeck, Germany · University of Southampton, Southampton, UK · Bonn-Rhein-Sieg University of Applied Sciences, Sankt Augustin, Germany

Abstract

Multi-agent reinforcement learning (MARL) has become a prominent research direction, particularly within collective robotics, where the ARGoS simulator is widely used to study emergent swarm behavior. Yet despite this growth, systematic support for reuse and reproducibility in MARL-based swarm robotics remains limited and insufficiently standardized: a single experiment is assembled from manually aligned simulator files, training scripts, configuration values, paths, seeds, and environment details, so reuse and reproduction stay fragile even within one research group. To address this gap, we introduce MARGoS, a lightweight, configuration-driven platform developed in this project that elevates these experiment-defining inputs to first-class, platform-managed artifacts. MARGoS enables reproducible experimentation through configuration hashing, end-to-end seed propagation, frozen (immutable) configurations, structured logging, environment fingerprinting, and exportable experiment bundles that facilitate sharing, replay, and verification of results across machines and over time. A five-dimension evaluation (architecture, efficiency, reproducibility, usability, and collaboration) supports our claim that managing experiment variation as a platform responsibility makes ARGoS+MARL experiments easier to reuse, reproduce, and hand off. Our code is available at https://github.com/tombackert/margos.

Figures & tables

Dimension	Question	Evidence used in this paper
RQ1 Architecture (enablement)	Can the platform execute the evaluation protocols?	Protocol enablement across RQ2-RQ5
RQ2 Efficiency	Does integration reduce repeated workflow effort?	Time and step reduction vs manual workflow
RQ3 Reproducibility	Can identical experiments be rerun reliably on the same hardware?	Reference run plus 20 automated reproductions
RQ4 Usability	Does the design reduce interaction complexity?	Heuristic audit and KLM analysis
RQ5 Collaboration	Can experiment bundles be handed off reliably to a prepared collaborator?	Export/import handoff trials

Table 1: Project goals and evaluation dimensions. Each dimension maps to a guiding question and to the evidence used to answer it in this paper; RQ1 is enablement-validated through protocol execution, RQ2-RQ5 are threshold-tested.

Panel of Figure 1 — Figure 1: MARGoS architecture. Control flow runs top-down from the five-command CLI through the five-stage pipeline (configuration $\to$ orchestration $\to$ logging $\to$ analysis $\to$ export/import); managed-artifact flow runs from each stage into the bottom row (frozen configs, runs and seeds, structured logs, reports, exportable bundles with environment fingerprints). The run command intentionally fans out because it spans configuration, orchestration, and logging rather than naming one stage. The ArgosToZoo bridge couples the orchestration stage to ARGoS via ZeroMQ.

Dimension	Primary metrics	Sample	Target
RQ2 Efficiency	Time-to-Complete; Steps-to-Complete	N=5 per condition	$\geq$ 50% reduction
RQ3 Reproducibility	Reproduce-Success-Rate; Result-Variance	N=20 automated runs	$\geq$ 90%
RQ4 Usability	Heuristic compliance; KLM reduction	35 criteria; 7 tasks; 1 structured audit	$\geq$ 80%; $\geq$ 50%
RQ5 Collaboration	Steps/time-to-share; first-run/reproduce time; handoff success	10 manual + 10 platform cycles; timed N=10/condition	$\geq$ 50%; $\geq$ 90% success

Table 2: Executed evaluation design: primary metrics, sample sizes, and predefined practical targets across the four empirically tested dimensions. RQ1 Architecture is omitted because it is validated by protocol enablement: the platform must execute the RQ2-RQ5 protocols end-to-end rather than pass a separate threshold test.

Dimension	Metric result	Threshold	Judgment
RQ2 Efficiency	Time -78.2%; steps -66.7%	$\geq$ 50% reduction	Supported
RQ3 Reproducibility	100% success (20/20); 0.0000 SD	$\geq$ 90% success	Strongly supported
RQ4 Usability	85.7% heuristic compliance; 56.9% KLM reduction	$\geq$ 80%; $\geq$ 50%	Supported
RQ5 Collaboration	100% handoff success; share time -65.3%; reproduce time -45.9%; first run +7.4% slower	$\geq$ 50% reductions; $\geq$ 90% success	Partially supported

Table 3: Measured outcomes across the four empirically tested dimensions. RQ2-RQ4 meet their predefined practical targets, while RQ5 is partially supported; jointly the four rows are the evidence bundle that platform-owned managed variation translates into measurable workflow effort reduction, deterministic self-reproducibility, analytical usability gains, and verified handoff at a first-run overhead.

Figure 2: Cross-dimension effort reductions on the executed project evaluation: every metric is plotted against the manual baseline at 0%, and the predefined

\geq

50% target appears as a dashed line. The chart is a visual comparison, not pooled inference: it juxtaposes workflow-effort gains on the RQ2 reference task, analytical KLM reductions on the RQ4 task set, and sharing-phase gains on RQ5, while showing the RQ5 reproduce-time metric narrowly below the 50% target.

Figure 3: Collaboration tradeoff (RQ5): the four timing metrics relative to the manual baseline at 0%. Large sharing-phase gains and reliable handoff success are set against the first-run overhead introduced by bundle validation and import checks. The figure shows the RQ5 partial verdict as a single tradeoff in the evidence bundle rather than as a missed target in isolation.

Explore similar work

MAGIC: Multi-Step Advantage-Gated Causal Influence for Multi-agent Reinforcement Learning
May 3, 2026Haohan Yu, Jinmiao Cong, Shengzhi Wang +2Multi-Agent Coordination Intrinsic Reward Methods
1Dalian University of Technology · 2Microsoft, Beijing, China
GzDRL: Reproducible and Scalable Deep Reinforcement Learning with Gazebo
Sep 3, 2026Amal Dev Haridevan, Junjie Kang, Jinjun ShanRobotic RL Reinforcement Learning
JaxMARL: Multi-Agent RL Environments and Algorithms in JAX
Nov 16, 2023Alexander Rutherford, Benjamin Ellis, Matteo Gallici +18RL Benchmarks Reinforcement Learning
University of Oxford · Universitat Politècnica de Catalunya · University College London +2

May 3, 2026, cs.MAJ/K move · Enter open · S save

MAGIC: Multi-Step Advantage-Gated Causal Influence for Multi-agent Reinforcement Learning

Haohan Yu, Jinmiao Cong, Shengzhi Wang, Lu Wang+1

1Dalian University of Technology · 2Microsoft, Beijing, China

Multi-Agent Coordination Intrinsic Reward Methods Reinforcement Learning Cooperative MARL Causal RL Reward Design for RL+1

Open paper AskPDF arXiv

A key challenge in multi-agent reinforcement learning (MARL) lies in designing learning signals that effectively promote coordination among agents. Designing such signals requires estimating how one agent's current action affects its teammates over future interaction steps. To address this, we introduce Multi-step Advantage-Gated Interventional Causal MARL (MAGIC), a framework that estimates multi-step action effects between agents and selectively converts them into intrinsic rewards. MAGIC uses counterfactual action interventions to compare teammate futures under factual and counterfactual branches, and introduces a gate based on advantage to direct exploration toward beneficial behaviors aligned with the task goal. Experiments on Multi-Agent Particle Environments (MPE) and StarCraft micromanagement benchmarks (SMAC and SMACv2) show that MAGIC consistently outperforms leading prior methods, with average relative final performance improvements of 26.9% and 10.1%, respectively.

Abstract

Explore similar work

MAGIC: Multi-Step Advantage-Gated Causal Influence for Multi-agent Reinforcement Learning

GzDRL: Reproducible and Scalable Deep Reinforcement Learning with Gazebo

JaxMARL: Multi-Agent RL Environments and Algorithms in JAX