Matching capabilities: SAS with its physical quantities.
Abstract
We present a cognitive mobile robot that acquires knowledge, and autonomously learns higher-level abstract capabilities based on play instincts, inspired by human behavior. To this end, we (i) model skills, (ii) model the robot’s sensor and actuator space based on elementary physical properties, and (iii) propose algorithms inspired by humans’ play instincts that allow the robot to autonomously learn the skills based on its sensor and actuator capabilities. We model general knowledge in the form of competencies (skills) of the mobile robot based on kinematic properties using physical quantities. Thus, by design, our approach has the potential to cover very generic application domains. To connect desired skills to the primitive capabilities of the robot’s sensors and actuators, it playfully explores the effects of its actions on its sensory input, thus autonomously learning relations and dependencies and eventually the desired skill. KnowRob is used for knowledge representation and reasoning, and the robot’s operation is based on ROS. In the experiments, we use a millirobot, sized 2 cm2, equipped with two wheels, motion, and distance sensors. We show that our cognitive mobile robot can successfully and autonomously learn elementary motion skills based on a playful exploration of its wheels and sensors.
Keywords
- artificial intelligence
- autonomous learning systems
- cognitive architecture
- reinforcement learning
- knowledge representation and reasoning
- resource-constrained systems
- low-energy mobile robots
1. Introduction
Our starting point is a robot with (a) a set of sensors and actuators, (b) tight resource limitations, (c) access to a database that captures general motion-related competencies (e.g. moving along a rectangle or navigating to a target location), and (d) built-in assumptions about physical laws and geometric relations. Our objective is to develop methods that allow the robot to autonomously learn competencies stored in the database.
Initially, the robot does not know the meaning and effect of its sensors and actuators (e.g. if an actuator controls a LED or a wheel). Therefore, the first activities are concerned with learning the meaning of its sensors and the effects of its actuators. Then, basic competencies from the knowledge base are acquired followed by increasingly complex competencies. A priori, the robot only has built-in knowledge of how to interface its sensors and actuators and basic assumptions about physical laws and geometric relations, but not what the sensors and actuators mean or how a specific motion can be accomplished.
Our long term goal is to provide the robot with general methods that allow the robot to work with any kind of sensors and actuators, in any kind of physical environment, and learning any kind of competence, provided it is possible at all (e.g. if the robot has only LEDs but no motors, it cannot learn to move).
We consider this a worthwhile vision because this approach to minimize prior knowledge and assumptions will facilitate very flexible systems that can work with any kind of sensors and actuators, in wheel-equipped or flying robots, on level plains, rocky or grassy surfaces, or even in wet environments. It will allow the use of accurate or inaccurate sensors and actuators, and to adapt to aging and wear-out effects. This approach is general because the only assumptions we make are the laws of kinematics and geometry, the availability of and access to sensors and actuators, the availability of learning methods (e.g., RL), and the availability of a database describing the skills to be learned.
While this is our vision, in this article, we make the further assumption that the robot knows the meaning of its sensors and operates in a two-dimensional plane. Inspired by the play instinct observed in humans and animals we propose exploratory, hierarchical learning. Simple and elementary tasks are tried out and learned first, followed by complex and composite tasks. This means the robot starts by asking if it can move at all, then it tries to learn elementary linear and angular motions, based upon which it studies moving along rectangles and similarly simple shapes. For each learning task, we use Reinforcement Learning (RL) as it matches well the exploratory nature of the robot’s setting. The learning tasks are identified based on entries from a knowledge database that describes the motion skills and the hierarchical relation between skills. Specifically, we use the KnowRob knowledge processing system [1], which is designed to provide autonomous robots with the knowledge base for performing motion and manipulation tasks.
In this paper, we propose and demonstrate the Skill Acquisition Method (SAM) for the case of a wheel-equipped tiny robot operating on a smooth, level plain; in future work, we will show that the same techniques generalize to other settings and environments. We evaluate our approach in a simulation environment for a two-wheeled and a four-wheeled mobile robot moving in a two-dimensional space. Experiments show that the system can learn and interpret its basic motion commands and derive complex motions, and finally, it succeeds in driving a rectangle (set of basic motion commands). Our contributions are summarized as follows:
i) We identify a minimal set of prior knowledge mandatory for learning basic movements.
ii) We propose a cognitive system behavior, the Playful Continuous Competence Acquisition (PCCA), that enables the learning and development of skills based on
a) the model of generic competencies (skills),
b) and the system’s Sensor and Actuator Space (SAS) grounded in elementary physical properties.
2. Related work
The use of knowledge representation and reasoning in robots has a long tradition, where the
In this context, KnowRob was specifically developed to equip autonomous robots with knowledge and methods (Knowledge Representation and Reasoning (KR&R)) to perform everyday manipulation tasks and to provide an infrastructure for cognitively enabled robots [1, 3, 4]. It represents one of the most advanced knowledge processing systems for robots, which has evolved even further with OPEN-EASE [5], which integrates KNOWROB2 [6], and aims to provide a remote knowledge and reasoning service that offers unprecedented access to the knowledge of autonomous robotic agents performing human-scale manipulation tasks. This seems promising for agents performing such rich human-scale manipulations but also places significant demands on the system’s resources, which is crucial for systems with limited resources. Therefore, we use KnowRob as the basis for knowledge processing and representation to take full advantage, but we target the approaches and methods that allow it to be deployed in such tiny systems.
A recent work dealing with the generalization of experience into abstract knowledge for novel situations, entitled Socio-Physical Model of Activities (SOMA [7]), consists of a comprehensive model for connecting physical and social entities that enable flexible execution by robotic agents. Since this representation seems essential, we use a similar approach, keeping our model flat in the first line due to resource-constraints. This limits the flexibility of the application (smaller knowledge base) but is crucial, and we aim for a reasonable trade-off. In this context, we also discuss a set of a small amount of prior knowledge.
RoboEarth has similar goals and approaches to our work [8]. Capabilities are also modeled, where we differ mainly in how they are used. We assume a set of general prior knowledge and basic methods to acquire skills, while their work accepts more complex algorithms to derive specific knowledge. Additionally, we further evaluate and improve skills to achieve continuous development.
Other works also deal with systems that learn semantically from different experiences, taking different approaches [9]. While learning relies on recorded experiences in semantic structures containing high-level representations. A key difference in our approach is that we generate skill-specific episodic knowledge through real-time learning methods, leading to knowledge abstraction at an earlier stage. To further leverage this, we define a set of prior knowledge that must be present to enable use in resource-constrained systems.
3. Skill acquisition
In the following, we present SAM, which starts from a set of general assumptions (knowledge and methods) to autonomously acquire and develop specific complex capabilities and aims for generic deployment in resource-constrained systems.
3.1 Overview
SAM (Figure 1) consists of various elements structured in layers. The bottom layer reflects the physical part, i.e., the robot and its environment. The layer above hosts the central computational agent, which abstracts the interface to the environment via the SAS. This general and generic interface is deliberately based on the fundamental physical properties of sensors and actuators. Thus, by definition, any environment can be integrated elegantly and efficiently as long as it follows the matching properties, defined in 3.4. Further, the agent has access to the knowledge base and reasoners. SAM follows a cognitive-behavioral architecture to autonomously learn skills using a KR&R methods combined with real-time learning from the physical environment.
3.1.1 Cognitive model
Cognitive models go beyond traditional behavioral models regarding what an entity (robot) knows, how that knowledge is acquired, and how it can be used. As a result, they are becoming increasingly popular in artificial intelligence. They are well suited for implementing highly autonomous systems that exhibit some intelligence and are expected to develop over time. There are several approaches to these models in the literature, particularly in robotics, which attempt to mimic the behavior of intelligent agents based on human cognition. Recent work on a generic form of this, such as the Socio-physical Model of Activities (SOMA) consists of a comprehensive model that combines physical and social entities and allows flexibility of execution by robotic agents through symbolic reasoning [7].
Since cognitive models, in the broader sense, represent complex processes and behavior, we focus our modeling on the core elements that we consider essential for our millirobot to acquire basic skills. In the medium term, we intend to adapt them to SOMA.
Figure 2 illustrates our proposed cognitive system model, divided into the physical and mental domains. The SAS depends on the physical properties of the robot. We divide the mental part into elementary capabilities and behavioral methods. A competence reflects knowledge of a particular skill acquired and applied through the PCCA behavioral method, while RL is used to learn a specific competence (e.g. motion commands). We will introduce and discuss these essential elements step by step in this article.
3.1.2 Use of KR&R and episodic memory
For each activity that SAM performs and observes (physical interaction, knowledge inference, learning, etc.), it generates skill-specific knowledge as episode memory and stores it with timestamps. Such episodic memory could include what the robot saw, reasoned, and did, how it did that, why, and what effects it caused [5]. It can be used for further conclusions and learning at any time. While the size and scope of the episodic memory directly relate to the resources required for the particular system. Many approaches attempt to collect a large amount of extensive detailed knowledge, which directly impacts computing time. This seems impractical for systems with limited resources. Therefore, we propose to keep episodic knowledge flat and small and to store only highly relevant information. In this context, we also consider a set of general prior knowledge that an intelligent system must have to learn and exhibit sufficient episodic memory for a given skill. We argue that these two facts are essential to consider for use in systems with tight resource limitations. Section 3.5.1 outlines an approach to a set of concrete prior knowledge and episodic knowledge developed by SAM, intended for use in resource-constrained systems.
Our long-term vision is that all relevant parts of the proposed SAM are hosted on such a system, e.g., a tiny millirobot powered by a micro-controller. We are aware of the challenges of migrating databases and logical reasoning to resource-constrained systems. As an intermediate step, we propose separating the acquisition and exploitation phases, where the system has access to KR&R in the first phase. Once the skill has been successfully acquired (sufficient episodic memory) to some degree, the system may be able to master it independently. Then it exploits the acquired skill with appropriate methods on the tiny millirobot. Whenever the system detects significant changes or decides to search for new capabilities, it contacts the database again. In this way, we can elaborate a similar knowledge acquisition behavior for resource-constrained systems compared to those with fewer constraints that host KR&R directly.
3.2 Competence
A central core element of SAM is competence, generally understood as mental property. The focus of this work is on the modeling of competencies that, when defined, lead to physical actions of the system through the SAS. However, competence in itself does not always have to be related to the physical facts of the system. It could also be a purely mental ability, such as spatial awareness, concentration, attention, reasoning, logic, and so forth. To model capabilities in an intelligent system, essential basic elements of those capabilities must be considered to grant an appropriate developmental progression. In a nutshell, a system should learn a skill independently and reason with appropriate knowledge about how good that skill is. Moreover, the evaluation of skills is of particular interest, used to continuously improve the respective skills. In this way, a cognitive system that also has an interest in developing itself further can become better over time.
In this context, two fundamental elements of competence have been attributed. These are (i)
Figure 3 illustrates the general concept of competence modeled in the knowledge base. The
3.3 Playful continuous competence acquisition
Another key core element is behavior, which ensures continuous development by learning new skills and further improving existing ones. Generally, a system that acquires specific skills should not consider them finally learned after the first success. Instead, the goal is to evaluate what has been learned and, if necessary, to develop further and improve it. In this way, a system can evolve autonomously and continuously adapt to certain changes in its environment. To this end, we consider the following key behavioral elements crucial: (a) the striving for new skills and (b) the continuous improvement of already learned skills.
Figure 4 illustrates our proposed PCCA method, focusing on knowledge acquisition and skill development. An interpretation of the learned skills in terms of possible application scenarios and their combinations in specific contexts, i.e., for which purpose skill could be used, is future work and not considered here. Further, to generalize the high-level system flow, a promising approach would be to model it directly in the knowledge base in tasks and actions. For that, KnowRob offers a promising approach that might also be applicable to our system [1, 7].
However, SAM’s high-level behavioral process is determined using the PCCA reasoner, directly queried by the computing agent. We define two different high-level-behavioral phases acting on the competence model properties (
3.3.1 Seeking new competencies
SAM searches for new capabilities based on the instances available in the knowledge base. Currently, these still need to be instantiated manually, with the long-term goal being to create them automatically. If one is present, the system uses the competence’s
3.3.2 Improving known competencies
The system decides whether a competence can still be improved based on the
3.3.3 Skill-specific learning methods
A specific competence is explored, learned, and exploited using appropriate learning methods (RL, supervised/unsupervised learning). These methods are competence specific and must be designed according to the particular skill. In principle, it is possible to integrate highly optimized learning algorithms for the respective functions. However, our goal is to use basic algorithms and execute them using general knowledge modeled in the knowledge base. In this way, we expect even more flexible usage, where only the primary parameters in the database need to be adjusted while the algorithm remains the same. When needed, the skill-specific learning method is triggered by the PCCA. In Section 3.5.4, we further discuss this approach and propose an RL basic algorithm that we extend with methods from KR&R to achieve generalization.
3.4 Sensor and actuator space
The sensor and actuator space (SAS) represents a generic interface to the robot environment, solely based on physical quantities. For example, consider an Inertial Measurement Unit (IMU), an odometry sensing unit as sensors, and two motors as actuators. SAM’s
Data | Description | Physical Quantity | Unit |
---|---|---|---|
Yaw rotation | Angle | ||
Horizontal cartesian coordinate | Length | m | |
Vertical cartesian coordinate | Length | m | |
Wheel torque | Torque | N m | |
Wheel torque | Torque | N m |
We assume that these interfaces abstract the robot-specific sensor data and actuator commands. For example, how the respective motor of the robot is controlled (using a motor controller that takes the acceleration properties into account) needs to be modeled robotic-system-dependent. In our case, the respective sensor data pre-processing layer and the actuator data post-processing layer take care of this.
This approach is generic, and we argue that the system initially does not need to know which actuators or sensors it is dealing with. A long-term goal is to employ appropriate methods and knowledge to identify and learn its capabilities. The Semantic Sensor Web follows this approach, annotating sensor data with various semantic metadata (including physical quantities) [10]. Further, there is promising work in automatic semantic knowledge acquisition for sensor data, which aims to annotate raw data with semantic knowledge [11]. Thus, our approach aims to leverage generic interfaces to integrate those methods seamlessly in future work.
However, the specific experimental setup is illustrated in Figure 5 for a two-wheeled mobile robot. It is equipped with two motors (for a 4-wheeled robot, extended by two additional motors), each driving a wheel, an inertial measurement unit (IMU), and an odometry sensing unit (obtained from the simulation environment) that is used to reduce the drift error of the IMU over time using a Kalman filter [12]. We are well aware of the challenges to the precision of these sensor measurements required for stable localization, which is extensively discussed in many publications [13, 14]. However, we do not further discuss this and assume that the problem is well understood. In conclusion, with this generic design, any robot environment can interact with SAM as long as the required physical
3.5 Motion skills
As mentioned earlier, this work focuses on modeling competencies that lead to physical actions of the system through SAS. Considering this fact and the physical characteristics of a wheeled robot, specifically the actuators in the form of wheels, potential movement possibilities can be assumed. For that, we consider basic movements, which in turn are subdivided into atomic and more complex movements. In a broader sense, for atomic actions, the robot is assumed to always be stationary, moving by applying torque to the actuators and stopping when it is removed. Such an atomic motion thus represents a sub-element of a more complex motion. It is not claimed that those movements are the most efficient in terms of smoothness and speed. However, they still allow the robot to approach all positions in a given space. Figure 6 illustrates a set of motion skills where atomic movements such as angular and linear movements ground complex movement patterns such as rectangles, cycles, or even more generally, a navigation path. The acquisition of these skills occurs in the same hierarchical manner that enhances the physical learning methods discussed in the next section.
3.5.1 Hierarchical knowledge acquisition
Let us first consider the knowledge we can gain about a movement, which we draw from a small set of prior knowledge. Assuming the system has not yet acquired any specific knowledge about motion, it has first to find out whether it can move at all with its given actuators:
3.5.2 Basic motions
For an atomic, basic motion, we refer to the basic kinematic and dynamic properties of a system, where kinematics describes the relationship between coordinates in motion space. Dynamics correlates the torque and force in each joint (wheels of the robot) with the acceleration of the joint and the velocity over time. When the wheels touch the ground, these forces act indirectly on the overall system and thus cause it to move. With the aid of the kinematic properties, inferences about this resulting motion can be drawn. Motion control for mobile robots is extensively covered in the literature. To navigate accurately, kinematic or dynamic models are used to generate accurate motion commands, considering all effects, including the resulting tracking error [15, 16, 17, 18, 19]. We are aware of the challenges of designing or even learning motion controls that lead to accurate robot movements. Thus, our work demonstrates the possibility of a generic approach to learning movements with general knowledge, even if the movements are still subject to certain errors. We will address minimizing this error by following the same general approach in future work.
However, based on universal laws of physics, we derive atomic base motions, illustrated in Figure 8. The robot’s position is represented by a vector with a pair of numerical coordinates
3.5.3 Kinematic reasoner
KnowRob [1] provides a kinematic reasoner, which we adapt to our SAM’s needs. It derives motion-specific competence knowledge based on general kinematic laws and is utilized during hierarchical knowledge acquisition level
A basic linear motion pattern is detected when the robot’s angle does not change during an action, but the distance does. Further, we can restrict it to a
For (ii), we define the following reasoner to argue about specific distances and angles used in level
These rules represent a general knowledge of the kinematic properties of a two-dimensional system, where we argue that SAM can be easily extended to three-dimensional systems by adding appropriate kinematic reasoners.
3.5.4 Skill specific learning methods
A specific competence is explored, learned, and exploited using appropriate skill-specific learning methods. Since SAM primarily focuses on acquiring skills that lead to physical actions, the respective atomic motion commands have to be learned in real-time by interacting with the environment. An appropriate learning procedure is required, whereas reinforcement learning methods achieve good results in this domain. The method dates back to the early 1990s when Q-learning was already used to learn specific, mostly robotic, tasks. However, many works solve various tasks with RL, whereby these are primarily designed in a context-specific, goal-directed manner and without explicit general prior knowledge, which significantly limits the learning of complex skills. We attempt to overcome this with our approach by using generally formulated prior knowledge to learn skill-specific, in our case, atomic motion commands.
In the following, we introduce our RL-based approach, which we extend with KR&R methods. In RL, an agent interacts with its environment over periods of discrete time steps
We chose a model-free approach for the specific RL algorithm based on a simple Q-Table RL method [20] for resource reasons. Keeping the required resources low seems to be the most intuitive first step for tackling our long-term vision, where all relevant parts of SAM, including RL, are hosted on a resource-constrained system. Q-Learning is a value-based method, where the Q-value is computed from the action-sate value function (Eq. (1)). It seeks to find the optimal Q-value for pairs of states and admissible actions. During exploration, the agent computes and stores them in a Table (Q-Table), where the Q-value indirectly represents the optimal policy
The Q-value function is defined according to the Bellman equation and notated as:
Where
The design of the reward function is formulated to learn specific basic motion distances and angles, while the angular and linear motion skills are learned separately and denoted as:
In principle, they each reflect a simple assumption: the closer a performed basic movement is to the desired distance, the higher the reward for that action. Thus, these rewards can also be considered a piece of specific general knowledge and are assessed by the RL Reward Reasoner. The resulting extended Q-Learning algorithm (Algorithm 1) follows a traditional flow, where the reward
1:
2: Observe initial state
3:
4: Select an random action action
5: Execute
6: Observe new state
7: Observe reward
8: Calculate Q-value
9: Update
10:
3.5.5 Prior knowledge and episodic memory
As discussed in Section 3.1.2, we propose to keep episodic memory flat and small and to store only highly relevant information. Further, we seek a set of general prior knowledge (semantic knowledge and general methods) that needs to be provided to learn and exhibit sufficient episodic memory for a given skill. We argue that these two facts are essential to consider for use in systems with limited resources. The following outlines how this might be addressed specifically in the case of SAM.
Figure 10 illustrates a set of prior assumptions, including semantic knowledge and general methods for acquiring motion-related episodic knowledge. The KR&R part might be provided by an edge device during the acquisition phase, while for the motion-specific learning, we deliberately propose Q-Table RL that requires few resources and thus can be hosted directly on the tiny millirobot. Further, we memorize only the motion commands learned by the RL with a timestamp and
4. Experiments
To evaluate the proposed SAM, we base our experiments on a simulation of a millirobot. Based on ROS, we use Gaezbo as a simulation environment, a Python ROS node for the computing agent, and KnowRob for the knowledge base and reasoning methods. We show the development phases
The primary experimental question is whether SAM can a) autonomously learn a motion skill based on a small set of prior knowledge, b) evaluate and continuously improve it, c) exhibit reasonably good time performance, and d) cover a generic application on various robot models.
4.1 Two-wheeled model
In the first experiments, a two-wheeled robot with dimensions of 2 cm
For the initial fundamental question,
At the next level
In the first attempt, we achieve relatively good results in an early phase, after only a few minutes. This is promising for use in resource-constrained systems, as it meets the resource requirements for migration mentioned earlier. The acquired competence knowledge is further used in level
4.2 Four-wheeled model
Further, we extend our experiments to a four-wheeled robot model as a first step to verify the general applicability of SAM. The basic assumptions and implementations of the robot model remain the same, except for two additional wheels. Due to these two further actuators, the search space increases, which leads to significant differences in the learning phase (level
The reason for this is the hierarchical learning approach, where the level above is abstracted from the level below in terms of performance. This gives us confidence that SAM is well suited for generalization. The last image of our experiments shows the development of the rectangular path by the four-wheeled robot with SAM, which was successfully mastered in 50 min (see Figure 15).
In summary, we have demonstrated with our experiments that SAM can learn autonomously complex motion skills based on a small set of prior knowledge and can further develop them with reasonable good time performance. We showed the first step for a generic application to various robot models by demonstrating the different wheel-based models.
5. Conclusions
In this article, we introduced SAM, which starts with a set of general prior knowledge and appropriate methods to autonomously acquire and develop specific complex skills. It combines methods of KR&R with methods of learning from the physical environment and aims to be applied in resource-constrained systems. We proposed a cognitive behavior (PCCA), which enables the continuous acquisition of skills, their evaluation, and the further development and adaptation of already learned skills. To this end, we modeled generic competencies using ontologies and formulated SAS based on elementary physical quantities to build a generic interface to the physical environment. Specifically, we demonstrated SAM based on motion skills learned through a general knowledge of kinematics laws and geometry. Further, we applied hierarchical knowledge acquisition with RL to acquire basic and more complex movements. We argue that this approach is general because the only assumptions we make are the laws of kinematics and geometry, the availability of and access to sensors and actuators, and the availability of a database describing the skills to be learned. Based on this generic knowledge, we demonstrated the acquisition of basic motion and a complex movement where the robot successfully moved along a rectangular path. To prove the generic approach, we evaluated it through experiments with a two-wheeled and four-wheeled millirobot. Where the acquisition performance in terms of resources delivers promising results for further deployment of the method in resource-constrained systems.
Thus, in the first step, we have demonstrated a cognitive system that develops more complex behaviors with a set of general prior knowledge and appropriate methods to function in arbitrary environments. In this work, we still assume that the robot knows the meaning of the actuators and sensors, although these do not necessarily have to be present a priori. In the next step, we want to remove this assumption. There is promising work in automatic semantic knowledge acquisition for sensor and actuator data that could help address this problem in a meaningful manner, which we will investigate further. Moreover, we will continue to develop an even more general approach, where an exhilarating challenge in this context could be the applicability of our method in a three-dimensional system. In addition, there are still limitations to the use of KR&R methods in resource-constrained systems, which we discussed in this work. Another medium-term goal is to study SAM in resource-constrained systems. Therefore, we will specifically address the transition to a real resource-constrained system in the form of a millirobot. In summary, our first results indicate that the use of SAM has an advantage for generic applicability, and we will continue to try to advance this approach.
Abbreviations
RL | Reinforcement Learning |
SAM | Skill Acquisition Method |
PCCA | Playful Continuous Competence Acquisition |
SAS | Sensor and Actuator Space |
KR&R | Knowledge Representation and Reasoning |
SOMA | Socio-physical Model of Activities |
IMU | Inertial Measurement Unit |
References
- 1.
Tenorth M, Beetz M. KNOWROB — knowledge processing for autonomous personal robots. In: 2009 IEEE/RSJ International Conference on Intelligent Robots and Systems. 2009. pp. 4261-4266. DOI: 10.1109/IROS.2009.5354602 - 2.
Nilsson NJ. Shakey the Robot. Technical Report 323, AI Center, SRI International. CA, USA: Menlo Park; 1984 - 3.
Tenorth M, Beetz M. KnowRob: A knowledge processing infrastructure for cognition-enabled robots. The International Journal of Robotics Research. 2013; 32 (5):566-590 - 4.
Tenorth M, Beetz M. Representations for robot knowledge in the KnowRob framework. Artificial Intelligence. 2017; 247 :151-169 - 5.
Beetz M, Tenorth M, Winkler J. Open-EASE. In: 2015 IEEE International Conference on Robotics and Automation (ICRA). 2015. pp. 1983-1990. DOI: 10.1109/ICRA.2015.7139458 - 6.
Beetz M, Beßler D, Haidu A, Pomarlan M, Bozcuoğlu AK, Bartels G. Know Rob 2.0 — A 2nd Generation Knowledge Processing Framework for Cognition-Enabled Robotic Agents. In: 2018 IEEE International Conference on Robotics and Automation (ICRA). 2018. pp. 512-519. DOI: 10.1109/ICRA.2018.8460964 - 7.
Beßler D, Porzel R, Pomarlan M, Vyas A, Höffner S, Beetz M, et al. Foundations of the socio-physical model of activities (SOMA) for autonomous robotic agents. Frontiers in Artificial Intelligence and Applications. Volume 344. arXiv preprint arXiv: 2011.11972. 202 - 8.
Riazuelo L, Tenorth M, et al. RoboEarth semantic mapping: A cloud enabled knowledge-based approach. IEEE Transactions on Automation Science and Engineering. 2015; 12 (2):432-443 - 9.
Wyatt JL, Aydemir A, et al. Self-understanding and self-extension: A systems and representational approach. IEEE Transactions on Autonomous Mental Development. 2010; 2 (4):282-303 - 10.
Sheth A, Henson C, Sahoo SS. Semantic sensor web. IEEE Internet Computing. 2008; 12 (4):78-83 - 11.
Ganz F, Barnaghi P, Carrez F. Automated semantic knowledge acquisition from sensor data. IEEE Systems Journal. 2016; 10 (3):1214-1225 Conference Name: IEEE Systems Journal - 12.
Moore T, Stouch D. A generalized extended Kalman filter implementation for the robot operating system. In: Menegatti E, Michael N, Berns K, Yamaguchi H, editors. Intelligent Autonomous Systems 13. Vol. 302. Cham: Springer International Publishing; 2016. pp. 335-348 - 13.
Alatise MB, Hancke GP. Pose estimation of a Mobile robot based on fusion of IMU data and vision data using an extended Kalman filter. Sensors. 2017; 17 (10):2164 - 14.
Durrant-Whyte H, Bailey T. Simultaneous localization and mapping: Part I. IEEE Robotics Automation Magazine. 2006; 13 (2):99-110 Conference Name: IEEE Robotics Automation Magazine - 15.
Fierro R, Lewis FL. Control of a nonholomic mobile robot: Backstepping kinematics into dynamics. Journal of Robotic Systems. 1997; 14 (3):149-163 - 16.
Morin P, Samson C. Motion control of wheeled Mobile robots. In: Siciliano B, Khatib O, editors. Springer Handbook of Robotics. Berlin, Heidelberg: Springer; 2008. pp. 799-826 - 17.
Martins FN, Celeste WC, Carelli R, Sarcinelli-Filho M, Bastos-Filho TF. An adaptive dynamic controller for autonomous mobile robot trajectory tracking. Control Engineering Practice. 2008; 16 (11):1354-1363 - 18.
Das T, Kar IN. Design and implementation of an adaptive fuzzy logic-based controller for wheeled mobile robots. IEEE Transactions on Control Systems Technology. 2006; 14 (3):501-510 Conference Name: IEEE Transactions on Control Systems Technology - 19.
Antonini P, Ippoliti G, Longhi S. Learning control of mobile robots using a multiprocessor system. Control Engineering Practice. 2006; 14 (11):1279-1295 - 20.
Christopher JCH. Watkins and Peter Dayan. Q-learning. Machine Learning. 1992; 8 (3):279-292