This research establishes a model of generative fabrication in which agent-based models imbue physical material with digital agency. This is a process in which real-time feedback is developed between industrial robots and multi-agent algorithms to explore the generative potential of the interaction of computational and material agency. This design research represents an inversion of material agency, from which two key concepts have emerged: parallelism, and stigmergic robotics. Rather than encoding material behavior within digital models, physical material takes on digital behaviors through an inversion of material agency. Parallelism describes a hybrid of digital and material behaviors through the closeness of their interaction. Stigmergic robotics collapses design and fabrication process-es into a single operation where the robot operates as an extension of the digital agent generating form through a series of design behaviors operat-ing directly on physical material.

STIGMERGIC ROBOTICS

This approach draws from the logic of multi-agent systems, and in particular stigmergic behavior that generates complex order through the accretion and re-organisation of matter. Stigmergic systems such as those that underlie the for-mation of termite mounds or the self-organisation of ant trails, operate through the indirect communication of agents. In these systems the agent (termite/ant) interacts directly with its environment (pheromone deposition). It is through the feedback of the agent altering the environment and the environment affecting the agent that complex order emerges. This self-organising logic provides the basis for a stigmergic fabrication approach (Snooks 2012). While these systems typi-cally involve many simple agents, there is no need for direct interaction between agents as the environment becomes the substrate of communication. Consequent-ly the same level of complexity can be achieved through the interaction of the environment and either a single agent or a large population of agents - assuming the number of operations is equivalent.

The capacity of industrial robots to run autonomously for extended periods of time allows for either a single, or low population of robots, to form complex stigmergic assemblages. In this paper we describe two strategies: a robot deposit-ing and interacting with a volatile material, and secondly two robots interacting with a volatile material - one depositing material while another removes material.

This approach to stigmergic robotics has arisen from our long-standing design research involving multi-agent algorithms and behavioral approaches to formation, and has further developed within what is an emerging discourse on real-time robotics and sensor feedback systems. This emerging domain encompasses the parametric updating of models based on sensor feedback to the generation of tool operations based on scanned matter. The original contribution our research offers is the real-time interaction of material and computational agency, where the formation of architecture is directly driven by their behavior, rather than an a priori parametric model or generative algorithm.

CONTROL SYSTEM

The technical workflow for stigmergic robotics involves a feedback between the robot, deposition and subtractive tools, material behavior, vision system and multi-agent algorithm. Our series of design experiments use two KukaAgilus KR10 R1100 SIXX robots controlled by a real-time server through KukaRSI, expanding foam and plastic extruded from end-effectors, Microsoft Kinect 2 and Structure IO sensors and a Java and Processing based algorithmic environment (See Fig. 3).

KUKA RSI

KUKA’s proprietary Robotic Sensor Interface (RSI) software allows an external computer to communicate directly with the robot controller via XML strings sent using a User Datagram Protocol (UDP) server over a local ethernet connection. RSI provides functions for making corrections to the robot position and digital IO signals, and sends information to the control pc detailing robot joint angles, global position, velocity and IO values in 4ms intervals. Any correction made to the position of the robot is executed within a single interval, limiting the ability of RSI to make substantial modifications to the existing position or motion of the robot without exceeding the limits of the joint motors. However, the pilot project below demonstrates that provided a stable connection can be maintained between the robot controller and the external PC, most of the functionality that would typically be found within a simple KRL toolpath program can be reliably execut-ed through the single RSI_MOVECORR() command by handling path planning and network communications on an external server application. This allows for non-linear design workflows whereby adjustment to robotic behaviours can be made by designers on the fly, in addition to facilitating the integration of KUKA industrial robots within more complex robotic systems that may incorporate dedicated computer vision processors, large sensor networks, or remote opera-tions.

SERVER APPLICATION

A C# application was developed that runs two UDP servers concurrently to han-dle communication with external machines (External Server) and the robot (RSI Server). The RSI Server communicates with the KUKA controller over a local ethernet connection and maintains data transfer using RSI XML formatting. The External Server handles sporadic communication with other control systems over an internet connection. Communication between the two server instances occurs through reference to thread safe dictionaries, allowing updates to the position, speed or IO signals of the robot to be passed to the KUKA controller by any pro-gram capable of sending XML strings over UDP. The RSI Server calculates ac-celeration and interpolated trajectories to a given target position and stores them as a list of correction commands that can be executed at 4ms intervals and inde-pendently of an external connection. Any update to the target position provided by the External Server will trigger a recalculation of the RSI Server path. A con-sequence of this server architecture is that responsibility for issuing robot com-mands is shifted from the KRL program to an external control system that is wholly independent of the procedural limitations of KRL. Although we demon-strate an approach that dynamically assigns robot tasks and behaviours within a multi agent simulation in Java, the C# application could be integrated with other platforms such as Grasshopper/Firefly or Python over UDP.

PUBLICATIONS

• Snooks, R. and Jahn, G. 2016, 'Closeness: On the relationship of multi-agent algorithms and robotic fabrication', in Proceedings of the 2016 Robotic Fabrication in Architecture, Art and Design, D. Reinhardt, J. Burry and R. Saunders (ed.), Springer, Switzerland, pp. 218-229

• Snooks, R. and Jahn, G. 2016, 'Stigmergic accretion', in Proceedings of the 2016 Robotic Fabrication in Architecture, Art and Design, D. Reinhardt, R. Saunders and J. Burry (ed.), Springer International Publishing, Switzerland, pp. 398-409