Flexible robotic production demands joint decisions on process progression, material routing, resource allocation, temporary cooperation, and simultaneous execution, because each decision influences the feasibility of the others. Under decentralized control, each robot only accesses bounded local information, yet system progress depends on shared resources, material states, and workspace compatibility—mirroring cooperative multi‑agent decision making under partial observability and resource contention.
This paper introduces AssemblyGrid v1, a reproducible benchmark for repeated multi‑robot production. It integrates explicit process progression, decentralized observations, material transfer, temporary multi‑robot coalitions, productive concurrency, and geometry‑dependent feasibility within a single task‑level formulation. The benchmark defines three workload families—Flow, Coalition, and Concurrency—each offering three scenario levels. Task success and evaluation metrics are independent of learning rewards or solution methods, allowing learning‑based and non‑learning approaches to be compared on the same production problem.
Evaluation consists of executable conformance checks, mechanism studies, and algorithmic experiments. Experiments employ a privileged centralized reference controller, structured decentralized controllers, and multi‑agent reinforcement learning (MARL) methods, namely IPPO, MAPPO, and QMIX. Results demonstrate productive execution under both centralized and decentralized control. MARL experiments further show that decentralized policies can learn effective production behavior from local observations and actions, confirming AssemblyGrid as a controlled benchmark for studying cooperative decision making in flexible robotic production.
Review: AssemblyGrid v1 consolidates the core challenges of multi‑robot production into a quantifiable benchmark, providing a unified platform for algorithm comparison and especially validating the feasibility of decentralized learning strategies.