To address the All the Agents Challenge (ATAC), we developed a solution for an Industry 4.0 use-case scenario that allows autonomous agents to reason on knowledge graphs, perceive, decide and act for reaching their goals, and coordinate the whole process in their Web of Things (WoT) environments. Our approach integrates the research works performed in the WoT, Semantic Web and the Multi-agent System (MAS) communities. Finally, We discuss the engineering properties of the proposed solution.
In this section, we present the approach for a multi-agent system (MAS) that can monitor and
control a process to achieve a goal with the inclusion of physical devices a.k.a ”Things”. In
our pursuit to address the main technological requirements of the ATAC challenge, we design
and develop our solution based on Hypermedia MAS [
The initializer is responsible for initializing the set of agents and the DomainArtifacts of
the process based on a domain specific language description (cf. marker 1 in Figure
1). As
the ThingArtifacts will only be created by the agents during run time, the initializer does not
handle their setup. The agents are the entities that can make decisions on the goals, the actions
that are executed on the system artifacts to achieve a goal and on the interactions among the
agents. The autonomous agents abstraction in our system is based on Belief-Desire-Intention
(BDI) agents [
The agents can use a set of controllable and observable artifacts to reach their goals (cf.
marker 2 in Figure 1). The artifacts can be physical entities (ThingArtifact) that can sense and
control the environment (e.g., devices in a manufacturing line) or domain entities
(DomainArtifact) that enrich the conceptual knowledge or accomplishment of agents and their goals (e.g.,
knowledge graph crawler, AI planner). The T h i n g A r t i f a c t s are described uniformly via the
W3C Things Description (WoT TD) model [
The agents use the D o m a i n A r t i f a c t KnowledgeGraphCrawler to navigate the hypermedia
environment and update their beliefs about the environment. This artifact discovers the entities
in the MAS and the interaction mechanism (e.g., HTTP) exposed by the resources. To enable
agents decide autonomously, all the resources (e.g., production line, Things, etc.,) and the
relations between them are described semantically in RDF using domain-specific vocabularies.
The knowledge graph uniquely identifies every thing instantiated using Internationalized
Resource Identifiers (IRIs). Thus, agents can perceive specifications of the real things during
run time to create the respective T h i n g A r t i f a c t s . To fulfill the purpose of a given use case the
agents’ desires can be defined either manually by an engineer or produced automatically using
an automated AI planner artifact [
To achieve the main goal of the use case, e.g. producing filled yogurt cups, the agents should have social ability and interact with one another. In this work, the agents can communicate with each other via messages (cf. marker 3 in Figure 1). Such an interaction always involves two agents exchanging one message with each other. Broader message exchange patterns like interactions with several agents or even broadcasts to all, require the agents to setup messages to all responsible others or use a D o m a i n A r t i f a c t like a pin board or similar. The messages themselves are use-case specific but can include: (1) message type, (2) sender’s identifier, (3) either an identifier on a property to interact with or an action to invoke, and (4) any content or reason part to transfer details about the property or action. Therewith, the agents’ interactions with both agents and artifacts can be modeled.
The application scenario that we use in this challenge is a simulated manufacturing product line in the Industry 4.0 community for filling and packaging yogurt 1. In this scenario, the factory is composed of diferent workshops which have diferent goals: conveying workshop, filling workshop, potting workshop and the packaging workshop. Each workshop includes diferent Things that can be used to sense the current state of the environment using sensors (e.g., presence sensor) and perform actuations on actuators (e.g., robotic arm). The factory also includes external suppliers which are responsible for providing empty yogurt cups, yogurt supplies and package providers for the fulfilment of an order.
In this scenario, the conveying workshop initially, loads the storage rack with yogurt cups, places an order for picking a cup from the shelves, picks up the cups using a Cartesian robot on X/Z axes, places the cups on a conveyor belt and finally the cups move to the head of the conveyor belt. In the filling workshop, when a cup is identified below the filler, subsequently the magnetic valve of the filler opens and the robot starts to follow the moving cup on the X axis. While the cup is moving, yogurt is poured into the cup. When the cup is filled with a certain amount of yogurt, the valve closes and the robot returns to its initial X coordinate while the container keeps moving towards the end of the conveyor. The main responsibility of the potting workshop is to grasp the filled yogurt pots and release them to the next workshop. For this, we use a Bosch APAS robot which moves in six diferent directions and can control grasping/releasing actuations via the built-in camera. The packaging workshop uses a Cartesian robot on the X/Y axes to pick up the cups form the conveyor belt and place them in an empty package. Once the package is filled with six cups, the Cartesian robot places the package on the next conveyor belt and subsequently fetches another empty package from the package bufer.
WAT was developed as an extension to the AI for Industry 4.0 summer school2. We demonstrate
a prototype of WAT within the above-mentioned use-case scenario implemented in Python. The
source code is available on GitHub3. Also, a video of our simulated demonstrator is accessible
on Youtube4. The demonstrator consists of six autonomous agents which correspond (1) to the
cup provider, (2) dairy product provider, (3) vl10 agent for controlling the conveying workshop,
(4) dx10 agent for controlling the filling workshop, (5) apas agent for controlling the robotic arm,
and (6) the xy10 agent for controlling the packaging workshop. The factory workspace contains
seven artifacts, where six of them are modelled as T h i n g A r t i f a c t that the agents can observe
and act on, and a L i n k e d D a t a F u S p i d e r representing the K n o w l e d g e G r a p h C r a w l e r by using Linked
Data-Fu [
Each agent and artifact is represented as a process and agents exchange messages via multiprocessing pipes for demonstration purposes. The multi-processing pipes can be exchange with any other FIFO messaging exchange technique. For the demonstrator, the following modules were used from the AI for Industry 4.0 summer school2 resources: • use-case scenario applied to WAT • graphical user interface of the factory for demonstration purposes • factory resources described in the knowledge graph for reasoning
JaCaMo [
One of the applications that integrates the above-mentioned approaches to provide a Linked Data interface for a simulated building scenario is the Building on Linked Data (BOLD)8. Similar to our approach, the agents can use hypermedia to browse the environment and discover Things via LDFU utility.
Schraudner et.al. [
In the following we discuss the proposed approach in terms of diferent engineering properties: Heterogeneity: The proposed system is manually described by a domain engineer according to diferent vocabularies such as SSN/SOSA, SAREF, etc., for decoupling on the semantic level and based on W3C WoT TD for decoupling from specific device APIs. The production rules and inference rules developed for crawling the knowledge graph are independent of the specific production line and the devices involved. An update on the production line does only require to revise the knowledge graph with the descriptions of the production line and a developer needs to provide the agents plan for reaching a specific goal.
Extensibility: The system is developed on top of WoT, which means that it can be easily extended with any Thing that exposes it’s capabilities with the WoT TD.
Modularity: The overall system is developed with diferent components for the agents and the diferent types of artifacts. This means that new agents and artifacts can be easily added or replaced without afecting the overall system.
Reusability: By relying on components, they can be reused in diferent use cases with new goals, as well as with diferent T h i n g A r t i f a c t s or D o m a i n A r t i f a c t s .
Scalability: The system is scalable, as it can spawn many agents and artifacts using processes. However, we have not done experiments to show the performance. This can be further improved by distributing agents and artifacts to diferent hosts with an own TCP server and enable the interactions to be transferred via the network. Therefore, the agents can also be integrated with the microservices architecture.
Dynamicity: The system supports dynamic behaviour as it can identify the diferent production line components, their properties, and the access method dynamically based on linked data. Also, the current state of the machines are always retrieved from sensor data and depending on them the system continually evolves and reacts to changes. For example, depending on the sensed amount of yogurt in the filling machine, new yogurt is ordered from a third party yogurt provider. The current solution’s dynamicity is limited in terms of unexpected behaviours such as device failure. In a production system, an agent’s goal relies on the Thing’s availability, while the WoT devices themselves are dynamic and could join or leave the network at arbitrary times during execution, making their availability unpredictable at runtime.