Publications / 2019 Proceedings of the 36th ISARC, Banff, Alberta, Canada

Automated Clash Resolution of Rebar Design in RC Joints using Multi-Agent Reinforcement Learning and BIM

Jiepeng Liu, Pengkun Liu, Liang Feng, Wenbo Wu and Hao Lan
Pages 921-928 (2019 Proceedings of the 36th ISARC, Banff, Alberta, Canada, ISBN 978-952-69524-0-6)

The design of rebar in reinforced concrete (RC) structures is a mandatory stage in building construction projects. Due to the large number and complicated arrangement rules of rebar in each design code, it is impractical, labor-intensive and error-prone for designers to avoid all clashes (i.e., collisions and congestion) manually or partial automation by using computer software. Therefore, the building information modeling (BIM) technology has been employed in the present architecture, engineering, and construction (ACE) industry for clash free rebar design. However, it is worth noting that most of existing BIM based approaches are using optimization algorithms for moving components, which can only be applied for regular shaped RC structures. In particular, shapes of rebar are fixed which means the optimized path of rebar cannot bend to avoid the obstacle in current studies. Furthermore, most of the existing studies cannot meet design constraints after avoiding clash, lack automatic and intelligent identification and resolution of rebar clash for complex RC joints and frame structures. Therefore, we present a framework towards automatic rebar design in RC frame without clashes via multi-agent reinforcement learning (MARL) system with BIM. In particular, by treating each rebar as an intelligence reinforcement learning (RL) agent, we propose to model the rebar design problem as a path-planning problem of multi-agent system. Next, by employing FALCON (A fusion architecture for learning, cognition, and navigation) with immediate evaluative feedback as the reinforcement learning engine, we design particular form of state, action, and rewards for reinforcement MARL for automatic rebar design. In addition, the design of rewards and some strategies in MARL are presented for build-ability constraints. Comprehensive experiments on one-story RC building frame have been conducted to evaluate the efficacy of the proposed framework. The obtained results confirmed that the proposed framework with MARL is effective and efficient.

Keywords: Building Information Modeling; Reinforcement Learning; Multi-agent; Rebar Design; Clash Resolution; RC Joints