Highlights
- •Service robots should have the ability to make decisions and schedule tasks.
- •Introduced a bioinspired hybrid cognitive architecture based on PDDL symbolic planning and state machines.
- •The proposed architecture is divided into two systems: deliberative and behavioral.
- •Each system of the architecture is composed of two layers.
Abstract
Any service robot should be able to make decisions and schedule tasks to reach predefined goals such as opening a door or assisting users at home. However, these processes are not single short-term tasks anymore and it is required to set long-term skills for establishing a control architecture that allows robots to perform daily tasks. This paper presents MERLIN2, a hybrid cognitive architecture based on symbolic planning and state machine decision-making systems that allows performing robot behaviors. The architecture can run in any robot running ROS 2, the latest version of the Robot Operative System. MERLIN2 is available at https://github.com/MERLIN2-ARCH/merlin2.
Keywords
Code metadata
Tabled
1
| Current code version | v1.0 |
| Permanent link to code/repository used for this code version | https://github.com/SoftwareImpacts/SIMPAC-2023-26 |
| Permanent link to Reproducible Capsule | https://codeocean.com/capsule/9822103/tree/v1 |
| Legal Code License | GPL-3.0 License |
| Code versioning system used | Git |
| Software code languages, tools, and services used | ROS 2, Python3 |
| Compilation requirements, operating environments & dependencies | Docker OR ROS 2, Python3 |
| If available Link to developer documentation/manual | https://merlin2.readthedocs.io/en/latest/ |
| Support email for questions | [email protected] |
1. Introduction
A social robot requires a considerable number of components working together for assisting a person in daily tasks. There are a number of challenges to overcome in a real scenario. The dynamics of environment such as humans walking randomly, long-term operations periods or managing multiple goals complicate decision making on the robot.
Under these circumstances, behavior generation in autonomous robots is an open problem associated with artificial intelligence systems. Several paradigms gave rise to what is known as control architectures [
[1]
], particularly the authors’ approach focus on those cognitively inspired. For instance, we can find deliberative architecture [[2]
], focused on planning approaches or subsumption architecture [[3]
] and reactive architectures [[4]
], which face quick behavior generation. These architectures and their combination have had more impact on the community: the three-layer architecture of Gat [[5]
], which organizes the decision-making component of the architecture in three different layers; and the hybrid architectures [[6]
], which merges deliberative and reactive concepts in order to fulfill short and long-term goals. As a result, it is possible to find in the research community several implementations of similar approaches, each one developed for its own purpose with issues for repeatability, making the picture of cognitive architectures [[7]
] hard to follow and to deploy in real robots.These days we can find specific approaches for previous paradigms developed in ROS [
[8]
]. From a deliberative perspective, planning approaches such as ROSPlan [[9]
] or PlanSys2 [[10]
] can be used to produce plans to achieve the high-level goals of the robot. These applications are based on the Planning Domain Definition Language (PDDL) [[11]
] that has been used to represent the environment of the robot using symbolic knowledge. However, these approaches do not give an easy way to modify the knowledge and do not provide persistence.From a reactive perspective, there are several solutions, for instance, SMACH the hierarchical state machine generator [
[12]
] or behavior trees [[13]
]. The former has alternatives such as FlexBee and the later has become the most extended paradigm in behavior generation in 2022. However, to integrate all these solutions in a single architecture can be difficult and for its own, some of them lack of functionalities such as visualization or symbolic knowledge manipulation.This work posed the software implementation of a cognitive architecture called MERLIN2 (evolution of MERLIN [
[14]
]) fully compatible with ROS 2, the standard de facto in software development. It provides a cognitive architecture framework that suits the hybrid architecture paradigm but it also includes generic architectural tools for managing symbolic knowledge and scheduling robot behaviors.
Fig. 1Bioinspiration of MERLIN2 components.
2. Architecture
MERLIN2, which comes from MachinEd Ros 2 pLannINg, is a software architecture that represents the paradigm of a hybrid cognitive architecture for robots composed of four layers [
[14]
]. It comprises the main software functionalities for generating robot behaviors: knowledge manipulation, for world model representation; deliberative decision making system, for accomplish missions and behavior generation for answering dynamic environments. This organization of MERLIN2 resemble human cognitive capacities.In nature, cognitive capacities such as short memory, perception or sensory-motor; are performed by some cerebrum lobes of the human brain [,
16
]. Organizing MERLIN2 components as cerebrum cognitive functions gave us the sketch presented in Fig. 1. The human cerebrum is composed of four structures called prefrontal, temporal, occipital and parietal [[17]
]. The prefrontal lobe has most of the modules from MERLIN2. It is in charge of memory, which is divided into long-term and short-term. The long-term memory, which represents the state of the environment, corresponds with the knowledge base managed by KANT. The short-term memory corresponds with the blackboards of the state machines of YASMIN. Besides, long-term memory is used in the planning process to choose the actions that the robot has to use to achieve its goals. These actions are implemented with YASMIN state machines and can also manage the knowledge of long-term memory. Finally, actions can use the navigation, text-to-speech and speech-to-text modules to interact with the environment and humans. The prefrontal lobe is also in charge of motor control, navigation, and speech production, text-to-speech; while the temporal lobe is in charge of the audio processing, speech-to-text.These cognitive functions are translated to the classic cognitive architecture depicted in Fig. 2. It presents in detail the MERLIN2 architecture technically divided into four layers classically associated with hybrid architectures and multi-layer cognitive proposals.
Fig. 2MERLIN2 architecture.
2.1 Mission layer
This layer is in charge of generating the high-level goals of the robots. These goals are related to robot duty, particular works that are internally or externally motivated, which means that are bred from the proprioceptive or exteroceptive stimulus (low battery status or human request).
The software component at this layer sets the goals of the robot in a given instant. That goal starts from an initial state made of current environment facts and sets the facts that the robots want to be true.
2.2 Planning layer
This layer is in charge of generating plans to achieve high-level goals previously created. It is associated with the classic Deliberative subsystem and requires formal representation of the environment, plan solvers, task dispatcher, and execution [
[2]
].MERLIN2 proposes a set of five components for this layer:
- •Knowledge Base: this component is the knowledge base of KANT, which corresponds with long-term memory. It stores the knowledge of the environment. There are two types of knowledge base depending on the storage type: a ROS 2 node that stores in process memory the knowledge and a MongoDB [[18]] database, which is more persistent.
- •PDDL Generator: this component is in charge of generating the PDDL text from the knowledge base.
- •Planner: this component uses the PDDL text to create plans. Several PDDL planners can be used. The available planners are POPF [[19]], SMTP [[20]], unified-planning [[21]] and VHPOP [[22]].
- •Plan Dispatcher: this component is in charge of executing the actions of the generated plan. It also has to cancel the current action if needed.
- •Executor: this component is a facade between this layer and the Mission Layer. It executed its component of this layer and cancel the Plan Dispatcher.
2.3 Executive layer
This layer is composed of the actions that the robot can carry out. When an action is instantiated, its associated knowledge is saved in the knowledge base. Besides, these actions are implemented as YASMIN state machines. As has been mentioned before, the blackboard of the state machines is the short-term memory, that stores the initial objects needed by the PDDL action and the partial solution of the state machine. Besides, YASMIN has a viewer that can be used to monitor the actions at a high level. Finally, new actions can be developed so that the robot can perform other tasks. As a result, more complex behaviors can be implemented.
2.4 Reactive layer
This layer provides the set of software components that bring bounded and specific applications that would be considered robot skills such as navigation (world exploration), text-to-speech (talking) or sound recognition (listening). This version of MERLIN2 includes three specific components that are suitable for validating the use of the cognitive architecture:
- •Navigation: this component corresponds with Nav2 [[23]], which is the navigation system of ROS 2.
- •Speech-to-text: this component is in charge of listening for speech and converting it to text.
- •Text-to-speech: this component is in charge of producing speech from a sentence.
3. Software impact
Cognitive architectures in robotics face the development and organization of software components to imitate the human mind. It is an active field of research where different software technologies and mechanisms can be found together. Some of them have been presented before, such as state machines and planning systems.
MERLIN2 cognitive architecture contains the elements that can be beneficial to research conducted in robotics competitions, in real and simulated environments, including but not limited to RoboCup or SciRoc. It allows the management of symbolic and sub-symbolic information through the use of KANT, generating Finite State Machine with YASMIN, and a mechanism for integrating robot tasks.
- 1.MERLIN2 aims to speed up the process of producing robot behaviors to face the daily short and long-term tasks of a service robot using ROS 2.
- 2.MERLIN2 combines four high-value tools for robotics developers: YASMIN, KANT, a Generic Planner System and a visual monitoring engine for mission and action tracking.
- 3.The four layers of MERLIN2 allow the generation of behaviors for performing complex robot tasks. The Mission Layer is supported by a mission engine which allows to the creation of an initial problem and the capability to create and execute goals as a consequence of developed actions. The Planning Layer manages symbolic task-solving. The Executive Layer has the actions, which are the result of associating and grouping skills based on two concepts: YASMIN Finite State Machine and (2) skill reuse. The Reactive Layer deploys the robot’s skills, adding a new one is a straightforward process. The current repository has examples with three basic robot skills: Navigation, STT and TTS.
- 4.MERLIN2 adds the feature for symbolic knowledge manipulation through KANT supported on DTO and DAO patterns. KANT provides two approaches: one for bringing knowledge persistence, using MongoDB; while the other uses in-memory storage. KANT engine makes it accessible to the developer to update/change the technology used.
- 5.MERLIN2 includes a generic planning system based on PDDL (Planning Layer). This planning system automatically extracts PDDL from the knowledge base, generates plans and executes the actions in a transparent manner for the developer. The solver motor includes simplified access to a set of PDDL solvers attending the version required.
The results enabled by the software individually and the whole architecture have been reported in the following academic publications:
- •Miguel Á. González-Santamarta, Francisco J. Rodríguez-Lera, Francisco Martín, Camino Fernández, and Vicente Matellán. Kant: A tool for grounding and knowledge management. In Josá Manuel Ferrán- dez Vicente, José Ramón Álvarez-Sánchez, Félix de la Paz López, and Hojjat Adeli, editors, Bio-inspired Systems and Applications: from Robotics to Ambient Intelligence, pages 452–461, Cham, 2022. Springer International Publishing. [[24]]
- González-Santamarta Miguel Á.
- Rodríguez-Lera Francisco J.
- Martín Francisco
- Fernández Camino
- Matellán Vicente
KANT: A tool for grounding and knowledge management.in: Ferrández Vicente José Manuel Álvarez-Sánchez José Ramón de la Paz López Félix Adeli Hojjat Bio-Inspired Systems and Applications: From Robotics to Ambient Intelligence. Springer International Publishing, Cham2022: 452-461 - •Miguel Á. González-Santamarta, Francisco J. Rodríguez-Lera, Vicente Matellán-Olivera, and Camino Fernández-Llamas. Yasmin: Yet another state machine. In Danilo Tardioli, Vicente Matellán, Guillermo Heredia, Manuel F. Silva, and Lino Marques, editors, ROBOT2022: Fifth Iberian Robotics Conference, pages 528–539, Cham, 2023. Springer International Publishing. [[25]]
- González-Santamarta Miguel Á.
- Rodríguez-Lera Francisco J.
- Matellán-Olivera Vicente
- Fernández-Llamas Camino
YASMIN: Yet another state machine.in: Tardioli Danilo Matellán Vicente Heredia Guillermo Silva Manuel F. Marques Lino ROBOT2022: Fifth Iberian Robotics Conference. Springer International Publishing, Cham2023: 528-539 - •Miguel Á. González-Santamarta, Francisco J. Rodríguez-Lera, Claudia Álvarez-Aparicio, Ángel M. Guerrero-Higueras, and Camino Fernández-Llamas. MERLIN a cognitive architecture for service robots. Applied Sciences, 10(17):5989, aug 2020. [[14]]
4. Limitations and further improvements
Finally, several future works are taken into account. MERLIN2 needs a perception module that enables more complex behaviors. Besides, it can be used to perform grounding and anchoring [
[26]
] improving the long-term and short-term memory of MERLIN2.Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
This work has been partially funded by an FPU fellowship provided by the Spanish Ministry of Universities (FPU21/01438) and the Grant PID2021-126592OB-C21 funded by MCIN/AEI/10.13039/501100011033 and by ERDF
A way of making Europe.
References
- A review of control architectures for autonomous navigation of mobile robots.Int. J. Phys. Sci. 2011; 6: 169-174
- Deliberation for autonomous robots: A survey.Artificial Intelligence. 2017; 247: 10-44
- New approaches to robotics.Science. 1991; 253: 1227-1232
- Experiences with an architecture for intelligent, reactive agents.J. Exp. Theoret. Artif. Intell. 1997; 9: 237-256
- On three-layer architectures.Artif. Intell. Mob. Robots. 1998; 195: 210
- AuRA: Principles and practice in review.J. Exp. Theoret. Artif. Intell. 1970; 9
- A survey of cognitive architectures in the past 20 years.IEEE Trans. Cybern. 2018; 48: 3280-3290
- ROS: an open-source robot operating system.in: ICRA Workshop on Open Source Software, vol. 3, no. 3.2. Kobe, Japan2009: 5
Michael Cashmore, Maria Fox, Derek Long, Daniele Magazzeni, Bram Ridder, Arnau Carrera, N. Palomeras, N. Hurtós, Marc Carreras, ROSplan: Planning in the robot operating system, in: Proceedings International Conference on Automated Planning and Scheduling, ICAPS, vol. 2015, 2015, pp. 333–341.
- PlanSys2: A planning system framework for ROS2.in: IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2021, Prague, Czech Republic, September 27 - October 1, 2021. IEEE, 2021
- PDDL2.1: An extension to PDDL for expressing temporal planning domains.J. Artif. Intell. Res. (JAIR). 2003; 20: 61-124
- The SMACH high-level executive.IEEE Robot. Autom. Mag. 2011; 17: 18-20
- On the implementation of behavior trees in robotics.IEEE Robot. Autom. Lett. 2021; 6: 5929-5936
- MERLIN a cognitive architecture for service robots.Appl. Sci. 2020; 10: 5989
- Human Brain Function.Elsevier, 2004
- An Introduction to Brain and Behavior.Worth New York, 2001
- Human Behavior and the Developing Brain.The Guilford Press, 1994
- MongoDB: The Definitive Guide: Powerful and Scalable Data Storage. The Expert’s Voice in Open Source. O’Reilly Media, Incorporated, 2019
Amanda Coles, Andrew Coles, Maria Fox, Derek Long, Forward-Chaining Partial-Order Planning, in: ICAPS 2010 - Proceedings of the 20th International Conference on Automated Planning and Scheduling, 2010, pp. 42–49.
Michael Cashmore, Maria Fox, Derek Long, Daniele Magazzeni, A compilation of the full PDDL+ language into SMT, in: Workshops at the Thirtieth AAAI Conference on Artificial Intelligence, 2016.
aiplan4eu, AIPLAN4EU/unified-planning: The AIPLAN4EU Unified Planning Library, GitHub.
- VHPOP: Versatile heuristic partial order planner.J. Artificial Intelligence Res. 2003; 20: 405-430
Steve Macenski, Francisco Martín, Ruffin White, Jonatan Ginés Clavero, The Marathon 2: A Navigation System, in: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS, 2020.
- KANT: A tool for grounding and knowledge management.in: Ferrández Vicente José Manuel Álvarez-Sánchez José Ramón de la Paz López Félix Adeli Hojjat Bio-Inspired Systems and Applications: From Robotics to Ambient Intelligence. Springer International Publishing, Cham2022: 452-461
- YASMIN: Yet another state machine.in: Tardioli Danilo Matellán Vicente Heredia Guillermo Silva Manuel F. Marques Lino ROBOT2022: Fifth Iberian Robotics Conference. Springer International Publishing, Cham2023: 528-539
- A short review of symbol grounding in robotic and intelligent systems.KI-Künstliche Intelligenz. 2013; 27: 129-136
Article info
Publication history
Published online: February 08, 2023
Accepted:
February 4,
2023
Received in revised form:
February 3,
2023
Received:
January 24,
2023
Footnotes
The code (and data) in this article has been certified as Reproducible by Code Ocean: (https://codeocean.com/). More information on the Reproducibility Badge Initiative is available at https://www.elsevier.com/physical-sciences-and-engineering/computer-science/journals.
Identification
Copyright
© 2023 The Author(s). Published by Elsevier B.V.
User license
Creative Commons Attribution (CC BY 4.0) | How you can reuse 
Elsevier's open access license policy
Creative Commons Attribution (CC BY 4.0)
Permitted
- Read, print & download
- Redistribute or republish the final article
- Text & data mine
- Translate the article
- Reuse portions or extracts from the article in other works
- Sell or re-use for commercial purposes
Elsevier's open access license policy
00014-3&id=gr1.jpg){kind=link}
00014-3&id=gr2.jpg){kind=link}