Industrial layout is decided early, by hand, and then lived with for a decade. Spatial Flow turns it into something you can search: the agent generates layouts, scores them against how the work actually moves, and hands back the trade-offs.
Context
IAAC · Barcelona
Year
2026
Stack
Reinforcement learning · Python · Spatial analysis
Brief
Spatial Flow is an AI-driven industrial layout agent developed to optimize the placement of machines, workstations, and circulation paths within production environments. The system combines computational design, spatial analysis, and reinforcement learning to generate and evaluate multiple layout configurations based on factors such as travel distance, operational sequence, accessibility, and collision avoidance. Through state graphs, stress testing, and benchmarking, the project explores how AI can support faster and more informed planning decisions while allowing designers to compare solutions and improve overall spatial and production efficiency.
01 · Context
Production layouts are set at the start of a project, from experience rather than measurement, and rarely revisited. The cost of a poor arrangement is paid every shift afterwards, in metres walked and sequences that double back on themselves.
02 · Agent
The agent places machines, workstations and circulation, then scores what it built against travel distance, operational sequence, accessibility and collisions. Each attempt informs the next, so the search concentrates where the gains are instead of sampling at random.
03 · Evaluation
A layout that scores well once proves nothing. Configurations are pushed through state graphs and stress tests so the comparison holds under changing volumes, and benchmarking keeps the agent honest against arrangements a planner would have drawn.
04 · Outcome
The designer is left with several defensible layouts and the reasoning behind each. That is the useful form: planning decisions get made faster because the trade-offs are explicit, not because a black box picked one.
