Robotic Reconfiguration with Deep Multi-Agent Reinforcement Learning

Our global population is estimated to increase to 11.2billion by the year 2100, requiring us to build 2 billion new homes over the next 80 years. The construction industry creates an estimated 33% of the world’s waste, and at least40% of the world’s carbon dioxide emissions (Miller 2021). The construction industry remains one of the least digitized and slowest to adopt disruptive technologies (Agarwal et al. 2016, Loosemore 2015). We continue constructing buildings organized in sheering layers and designed with linear building life cycles eventually ending in demolition (GlobalData n.d. 2018, Ngwepe and Aigbavboa 2015, Brand 1995). To address the unprecedented challenges of the global climate and housing crises requires radically changing the way we conceive, plan, and construct buildings, from static continuous objects to adaptative eco-systems of reconfigurable parts.

Living systems in nature demonstrate extraordinary scalable efficiencies in adaptive construction with simple, flexible parts made from sustainable materials. For example, nomadic ant colonies face extreme pressure to generate foraging routes, moving massive numbers of ants each day, yet through their simple parts and local rules they have shown rapid efficiency in constructing adaptive “living bridges” through the linkage of their bodies. Ants modulate their behavior in response to locally changing environments to adapt to dynamic traffic conditions, recover from damage, and disassemble when underused (Graham et al. 2017).

Inspired by robust natural construction systems, new approaches to construction with teams of robots have become active areas of interdisciplinary research highlighting opportunities for safe, sustainable, and efficient building construction. Collective robotic construction specifically concerns embodied, autonomous, multirobot systems that modify a shared environment according to high-level, user-specified goals integrating architectural design, the construction process, mechanisms, and control(Peterson et al. 2019). These systems typically involve machines that are codesigned with the architectural systems they construct, enabling them to be more adaptive, scalable, and reusable while operating in dynamic environments (Leder et al. 2022, Silver 2017, Lindsey et al. 2011, Kayser et al. 2018, Jenett et al. 2019, Terada and Murata 2008, Napp et al. 2012).

To demonstrate the opportunities inspired by natural systems, CRC systems must be developed with artificial intelligence for collaborative and adaptive construction, which has yet to be explored. Autonomous collaborative robotic reconfiguration is a robotic material system with an adaptive lifecycle trained with DMARL for collaborative reconfiguration. Autonomous collaborative robotic reconfiguration is implemented through three interrelated components codesigned in relation to each other: 1) a reconfigurable robotic material system; 2)a cyber-physical simulation, sensing, and control system; and 3) a framework for collaborative robotic intelligence with DMARL. The integration of the CRC system with bidirectional cyber-physical control and collaborative intelligence enables us to project operating as a scalable and adaptive architectural eco-system.