- Information
- AI Chat
Seminar report of ECE final year
Electronic and communication (ECE)
Visvesvaraya Technological University
Recommended for you
Preview text
Chapter 1
INTRODUCTION
1 Machine learning
Machine learning (ML) is the study of computer algorithms that improve automatically
through experience and by the use of data. It is seen as a part of artificial intelligence. Machine learning algorithms build a model based on sample data, known as "training data", in order to make predictions or decisions without being explicitly programmed to do so. Machine learning algorithms are used in a wide variety of applications, such as in medicine, email filtering, and computer vision, where it is difficult or unfeasible to develop conventional algorithms to perform the needed tasks.
A subset of machine learning is closely related to computational statistics, which focuses on making predictions using computers; but not all machine learning is statistical learning. The study of mathematical optimization delivers methods, theory and application domains to the field of machine learning. Data mining is a related field of study, focusing on exploratory data analysis through unsupervised learning. In its application across business problems, machine learning is also referred to as predictive analytics.
1 Robotics
Robotics is an interdisciplinary field that integrates computer science and engineering. Robotics involves design, construction, operation, and use of robots. The goal of robotics is to design machines that can help and assist humans.
Robotics develops machines that can substitute for humans and replicate human actions. Robots can be used in many situations and for many purposes, but today many are used in dangerous environments (including inspection of radioactive materials, bomb detection and deactivation), manufacturing processes, or where humans cannot survive (e. in space, underwater, in high heat, and clean up and containment of hazardous materials and radiation). Robots can take on any form but some are made to resemble humans in appearance. This is said to help in the acceptance of a robot in certain replicative behaviours usually performed by people. Such robots attempt to replicate walking, lifting, speech, cognition, or any other human activity.
1 Machine Learning in Robotics
Learning is a major part of the research in Robotics. Machine learning algorithms in robotics in particular, are being used to tackle learning tasks where large quantities of datasets are available which enable Robots to effectively teach themselves accordingly. Yet application of machine learning in Robotics which highly contributes to Robot learning is vast and yet progressing in a fast pace. Robot vision, Robot navigation, field Robotics, humanoid Robotics, legged locomotion, off-road rough-terrain mobile Robot navigation, modelling vehicle dynamics, medical and surgery Robotics are few of the areas within Robotics for which utilizing machine learning technologies has become popular.
It is therefore clearly evidenced that machine learning has in recent years become an essential part of Robotics. And this has been in fact a response to the frustration with the problems for which it has been proven difficult to conventional coding solutions. For instance in a variety of Robotics platforms such as humanoid robotics, and legged locomotion, the imitation learning techniques, and inverse optimal control methods play an increasingly important role. In such areas, for instance, the alternative approach of programming-by- demonstration is utilized where Robot behaviours are created by involvement of expert demonstration.
In addition the supervised learning techniques have become norm within field robotics in rough terrain. Furthermore self-supervised learning is applied to generate training examples for self-improvement enabling Robots to effectively teach themselves. A number of the most notable machine learning technologies utilized in Robotics realm includes; reinforcement learning, supervised learning, self-supervised learning, multi-agent learning, autonomous science, machine learning techniques for big data, imitation learning techniques, Robot programming by demonstration, and multi-agent learning.
Chapter 3
LEARNING AND OPTIMIZATION IN ROBOTICS
Machine learning as a sub-field of computer science has evolved from the study of pattern recognition and computational learning theory. Machine learning is considered as a field of study in artificial intelligence that gives computers the ability to learn from data. To do so machine learning explores the development of models that can predict and learn from an available dataset. Such models operate with the aid of algorithms capable of making data-driven predictions rather than following explicit codes. Consequently machine learning is often used in a range of problems where designing precise algorithms is not practical. In this sense machine learning can replace the human expertise in information treatment. To doing so machine learning provides the algorithmic tools for dealing with datasets and providing predictions. In fact machine learning tends to imitate human skills, which in most cases, act exceptional in identifying satisfactory solutions by theoretical or experience-based considerations.
Figure 3 Learning agile and dynamic motor skills for robot.
The intersection research area of optimization and machine learning has recently engaged leading scientists. Machine learning has made benefit from optimization and on the other hand machine learning contributed to optimization as well. Today machine learning is seen as an exceptional replacement for human expertise in information manipulation. In addition machine learning has the proven ability to simplify optimization functions. Optimization on the other hand is the source of immense power for automatically improving decisions. However in real- life applications, including Robotics, optimization has not had the chance to be used to its full
potential. This has been often due to the absence, complexity, or inefficient optimization functions of the complicated problem at hand. Yet in such cases machine learning has shown the ability of modeling whole or part of the optimization functions on the basis of the availability of a reliable dataset. A number of case studies concerning Robotics problems have been surveyed in literature, where machine learning technologies simplify complicated optimization functions.
Nevertheless the long-term vision for Robot learning would be the development of a fully automated system with self-service usage. To reach this goal the novel idea of integration of machine learning and optimization aims at simplifying the whole learning process by automating the decision-making tasks in an effective manner without requiring a costly learning curve for the final-user. In this context the learning process is seen as a byproduct of an automated optimal decision. Learning from the available dataset integrated with optimization can be applied to a wide range of complex, dynamic, and stochastic problems. Such integration has been reported exceptional in increasing the automation level by putting more power at the hands of final-user. Final-user should however specify dataset, desired outputs and CPU time. CPU time is to be set to put a limitation on optimization algorithms’ run-time which can be referred as learning time.
The novel integration of machine learning and optimization has already been used in solving numerous complex cases. Decision-making in complex geometrical problems, patient’s diagnosis problem and healthcare decision-making, multi-objective optimization problems wireless access point optimization, mobile Robot navigation, business intelligence and business decision-making models, automated decision-making bioinformatics and big data are few examples. Considering these examples, it is observed that once a combination of right machine learning technologies and optimization algorithms designed, suitable for the problem at hand, further algorithm selection, adaptation, and integration, are done in an automated way, and a complete solution for learning is delivered to the final user. As the result the real-life decision- making tasks with the diverse, stochastic, and dynamic nature can be handled smoothly and continuously. Within the framework of such integration, it is expected that algorithm selection and adaptation are done in an automated way, and a comprehensive solution is delivered to the final user.
control and decision-making subsystems to perform the task or mission we want with reasonable guarantees of safety.
4 Data driven design
When models cannot be found, or desired system behaviours are difficult to specify, one may decide to use alternative methods based on data collection as shown in Fig.4. However, when this approach is implemented with machine learning, accidents may happen when the data is noisy or if the system needs to explore new solutions. Therefore, the acceptable risk level needs to be taken into account when designing a data-driven system. An example involving human behaviour may help explaining this idea.
Figure 4 Data driven design
Data-driven programming is similar to event-driven programming, in that both are structured as pattern matching and resulting processing, and are usually implemented by a main loop, though they are typically applied to different domains. The condition/action model is also similar to aspect-oriented programming, where when a join point (condition) is reached, a pointcut (action) is executed. A similar paradigm is used in some tracing frameworks such as DTrace, where one lists probes (instrumentation points) and associated actions, which execute when the condition is satisfied.
4 Hybrid system
A hybrid system is a dynamical system that exhibits both continuous and discrete dynamic behaviour a system that can both flow (described by a differential equation) and jump (described by a state machine or automaton). Often, the term "hybrid dynamical system" is used, to distinguish over hybrid systems such as those that combine neural nets and fuzzy logic, or electrical and mechanical drivelines. A hybrid system has the benefit of encompassing a larger class of systems within its structure, allowing for more flexibility in modeling dynamic phenomena.
Figure4 – Hybrid system
The state of a hybrid system is defined by the values of the continuous variables and a discrete mode. The state changes either continuously, according to a flow condition, or discretely according to a control graph. Continuous flow is permitted as long as so- called invariants hold, while discrete transitions can occur as soon as given jump conditions are satisfied. Discrete transitions may be associated with events.
Figure 5 – Robot with common sensors
5 Deep Learning Artificial Neural Network
Data analytics and machine learning, can be used for extraction of features and/or data fusion. One such method is Convolution Neural Networks (CNN) with deep structures, popularly known as Deep Learning Artificial Neural Networks (DLANN). These networks have multiple layers of different types (convolution, pooling, etc.) in sequence, hence the use of the term deep (Fig. 5). However, for these methods to be useful, a large number of data points need to be available and the algorithms cannot usually be run in real-time.
Figure 5 – Deep Learning Artificial Neural Network Some improvements may be added to the data-driven machine learning approach by adding the model the robot to the sensing system. In this case, the output of the DLANN could
be constrained by the model of the robot such that the states could be better estimated. In this way, spurious results could be filtered out of the measurements and sensing improved.
5 Control
In the context of this article, a controller may be defined a device that is used to drive a robot to a desired state, usually a position and/or velocity reference. If we know the model of the robot, we can derive controllers to stabilize the system as well as to present a desired response (move faster or slower) with very small (or zero) error. When we do not know the model of the robot or when the environment interferes with its behaviour, we can use a system identification approach to find the parameters that define the motion of the vehicle. Several system identification approaches are commonly used for this task: from simple (least squares approximations) to complex ones. Machine learning approaches can also be used, the most of which are Fuzzy Systems and, again, Neural Networks. Neural Network approaches are usually referred to as “black-box” approaches, for the parameters do not have any physical meaning as parameters of a system. With CNN, however, one can introduce meaning to the parameters, as, in determining the model of a robot, one usually ends up with a convolution. Recent approaches use CNN to find Finite Impulse Response (FIR) or Infinite Impulse Response (IIR) linear models for robots and UAVs. The same approach can be used to find more sophisticated models such as Volterra series and Generalized Orthonormal Basis Functions (GOBF).
Fuzzy Systems, on the other hand, can be used in real-time. Fuzzy Systems are used to model systems using what is called linguistic variables. In this way, the system can be approached by using normal language such as near or far. By allowing the definition of the terms to change based on feedback from the environment, controllers may change over time.
Figure 5 – Fuzzy linguistic variable for distance.
5 Decision-making
A controller as discussed in the previous section is a decision-making device. As depicted in Fig. 7, a controller takes a reference command and, by measuring the outputs of the robot, provides commands that guarantee that the robot execute the desired task.
Figure 5 – Block diagram for the control of a robot.
Decision-making to mean the high level choice of strategies to be made by a robot. The main distinction is that the choice of strategies is not easily specified, since the desired outcome of the system is not clear, whereas the performance of the controller is well-stipulated in terms of transient and steady-state parameters. In a very broad sense, one may say that the functions of decision-making and control in robotic systems are analogous to the human nervous system.
The decision-making system’s function would be similar to the function of the central nervous system (CNS) whereas the controllers’ functions would be similar to the peripheral nervous system (PNS). The first is responsible for the conscious decisions while the second is responsible for the automatic response. Both are extremely important and in conjunction provide a smooth behaviour for the whole system. For designers or robotic systems the question then becomes how these strategies may be defined. Traditionally, one of the most popular ways to derive strategies for robots is through the use of Markov Decision Processes (MDPs) in several different flavours (POMDP, MOMDP, etc.). But in order to apply this type of method, a model of the environment including especially its probability distribution needs to be available. When the environment’s
probability distribution or the cost function that determines the behaviour of the robot are not readily accessible, it is complicated to use MDP-based solutions. In this case, another method that can automatically and implicitly extract the probability distribution may be successfully used. For example, reinforcement learning has been used to approximate the probability distribution of unknown environments. Also, automated derivation of rewards and cost functions have been proposed in the context of multiple agent environments. Both solutions depend on the availability of data, but they do not necessarily rely upon reacquired data. Therefore, these methods may be executed in real- time.
It must be noted that sometimes the design of machine learning solutions is not straightforward. One common criticism of these methods is that they only displace the complexity of the design from one subsystem (the model of the environment) to another (the machine learning algorithm). A rule of thumb to measure the effectiveness of any approach may be the number of parameters necessary to define the solutions. We always aim to have a smaller number of parameters to tune. It can be shown that in several applications, machine learning (or other big data approaches) is successful in satisfying this criterion.
5 Recognition Technology
Recognition technology is to be detected even if there are variations of
position, orientation, scale, partial occlusion and environment variations as
intensity. Object detection is the key to other machine vision functions such as
building 3D scene, getting additional information of the object (like face details)
and tracking its motion using video successive frames. With machine learning
the robot is able to recognize spoken commands to move correctly to give a
direction to robot, first the voice command is sent to the computer using a
microphone the computer recognize the command by speech recognition
system.
5 The Basic Robot Loop
Figure 5 Basic Operation of a Robot.
➢ Remote Control Interface/ Command
Remote control is about controlling a robot from a distance, either with or without a wire. Remote control methods can be split into two categories: wireless and wired. Wired remote
control or tethered control can be right way to interface a computer with a stationary robot.
➢ Path Planning Module
Path planning module is control of PC to a Robot. Path planning in order to compute the path the robot should follow to ensure safe navigation and exicute the desired mission. It finds a sequence of valid configuration that moves the object from the source to destination.
➢ Guidance and Control
Guidance and control system responsible for ensuring vehicle stability and reference object tracking as well as disturbance rejection. Here to investigate and develop advanced control system for guidance and positioning of robot with flexible links and joints.
➢ Robot real Motion
Real life robot expressing its dynamic behaviour in response to the control inputs and external disturbances. A motion control system is initiate and control the movement of a load to perform work. It is capable of precise speed position, and torque control such systems are typically comprised of basic components like controller, the drive or amplifier and motor. The controller plans the path.
➢ Perception and State Estimation
State estimation for robotics provides a very timely over view of state estimation in robotics. It contains some functions like – ➢ Inertial Navigation system: An inertial navigation system for a mobile robot. Error models for the inertial sensors are generate and included in an external kalman filter (it is an Algorithm that uses a series of measurement) for estimating the position and orientation of a moving robot. ➢ Localization and mapping: In robotics, simultaneous localization and mapping is the computational problem of constructing or updating a map of an unknown environment simultaneously keeping track of a location within it. ➢ Sensor Fusion: Sensor fusion is the process of combining sensory data or data derived from disparate source. Sensor fusion have the ability to bring together inputs from multiple radars and cameras from a single model or image. ➢ Sematic Understanding: Machine learning can allow robot t acquire complex skills, such as grasping & opening doors learning these skills required to understand semantic concepts to get them to follow simple command.
6 DISADVANTAGES
1. Higher Cost & Continues supply power
This is one of the disadvantages of robotics. Robots consume a lot of power to function. Robots need to be maintained continuously to keep them in good condition.
2. Unemployment
If robotics comes into trend, then many skillful workers would also lose their jobs and would be on roads, which is one of the disadvantages of robotics. Many daily wage workers would lose their jobs, which are actually the bread and butter of their families.
3. Time and Resources
ML needs enough time to let the algorithms learn and develop enough to fulfill their purpose with a considerable amount of accuracy and relevancy. It also needs massive resources to function.
6 APPLICATIONS
1. Assistive and medical technologies
An assistive robot is a device that can brain, process information, and execute actions that can help people with disabilities & seniors. And smart assistive technologies also exist for ordinary people or users like driver assistance tools. Movement robots give you a therapeutic or diagnostic benefit.
2. Computer Vision
Robots seeing involves more than just computer algorithms. Robot vision is very closely linked to machine vision, which can be given credit for the emergence of robot guidance and automatic inspection systems.
3. Self-Supervised Learning
Self-supervised learning approaches enable robots to generate their own training examples in order to improve performance; this includes using a priori training and data captured close
range to interpret “long-range ambiguous sensor data.”
Seminar report of ECE final year
Course: Electronic and communication (ECE)
University: Visvesvaraya Technological University
This is a preview
Access to all documents
Get Unlimited Downloads
Improve your grades
