PPO-based mobile robotic swarm for spatially adaptive architectural roof interface: a proof-of-concept

Purpose As an extension of the self-assembly architectural design, the purpose of this study is to present a proof-of-concept validation of whether decentralized deep reinforcement learning (DRL) can enable self-organizing collective behavior of robotic agents toward a futuristic adaptive building roof interface. Design/methodology/approach A physical mock-up comprising a 2.4 × 1.52 m plexiglass roof platform and 15 identical robotic balls was developed. A ball integrates a microcontroller, DC motors, batteries and a dual-wheel mechanism within a spherical casing (d = 140 mm). A proximal policy optimization (PPO)-based controller was implemented to train the ball agents to move toward circular target zones (R = 0.25 m) while reducing collisions and maintaining spatial dispersion for shading. Simulation and field-validation experiments were conducted with the full agent configuration. Performance was evaluated using target coverage ratio, reaction time, travel distance and collision count/frequency. Findings The agentic system achieved 35%–55% target coverage within about ∼5 min, outperforming Greedy and artificial potential field baselines in both goal-reaching time and path length. The results demonstrate the feasibility of linking artificial swarm intelligence with spatial reconfiguration in a built-environment context. Research limitations/implications The study is conducted at a proof-of-concept scale, using a small experimental mock-up and lightweight mobile robots. Originality/value To the best of the authors’ knowledge, this study presents one of the first validations of a decentralized DRL-based swarm robotic system for an architectural roof interface, demonstrating the feasibility of autonomous spatial redistribution for target-zone shading without centralized coordination.