ABSTRACT.
In our everyday lives, we interact with agents like personal computers, search engines, cars, etc., and reveal many of our personal choices, biases, and preferences. Improving agents by analyzing their interactions with humans is an active area of research. However, the algorithmic and systems challenges involved in interactive learning varies with the application. For example, there has been significant advances in learning from humans for information retrieval tasks such as recommendation systems [82, 126], ranking documents [102, 105, 199], etc. This is primarily due to the availability of scalable and easy-to-use interaction mechanisms (like search engines) that reap massive logs of interaction data. As things stand in robotics, interactive learning is in its early stages due to the involved systems challenges, data scarcity, scaling difficulties, and the multi-modal nature of robotics problems. Nonetheless, learning from humans is a necessary step to unlock natural human-robot interactions with applications ranging from household robots to autonomous cars. Interactive human-robot learning presents new opportunities for developing novel interaction mechanisms, algorithms for sensory-fusion, and learning preferences grounded in the physical-world. In this dissertation we study algorithms and applications of interactive learning between humans and robots. Figure 1.1 demonstrates two examples of interactions: one between a human and a household robot (Baxter) and another between a human driver and the car. In order for the agents (robot or car) to learn from such interactions, we address questions like: 1 0.1 0.2 0.7 Face Camera Road Camera Figure 1.1: Robot and human interactions • How do we design interaction mechanisms that are easy to use by non-expert users? • How do we build learning algorithms that can learn from weak human signals? • How do we model the rich spatio-temporal interactions using deep neural networks? 1.1 Robots and humans: Interactive learning In order for robots to operate autonomously, they need to learn many concepts about our physical world. The skills they should acquire ranges from perception to performing actions in the human environment. While several skills can be learned from large annotated databases (e.g. object detection using Imagenet [46]), many others require the robot to observe and interact with humans. In this dissertation we focus on skills that robots learn by interacting with humans. This includes learning context dependent actions, understanding and anticipating human behaviour, and generating desirable motions to fulfill tasks. 2 The interaction mechanism between the human and the robot is also an important aspect to consider. First, the interaction being analyzed should vary based on the application and the nature of skill that the robot wants to learn. Second, it should also take into account the sub-optimality of human signals. In this dissertation we address several human-robot interactive learning problems and propose mechanisms for eliciting human signals. We focus on learning from non-expert users and build interaction mechanisms that are easy-to-use – often as easy as a clicking a mouse – to widely elicit user interaction. Figure 1.1 shows some examples of robots interacting with humans. When interacting with robotic manipulators with high degrees-of-freedom, it becomes challenging for non-expert users to provide optimal kinesthetic demonstrations. In chapter 2 & 3, we consider such robotic manipulation and navigation tasks and learn user preferences over robot trajectories. We propose multiple easy-to-elicit feedback mechanisms and develop algorithms to learn from sub-optimal user feedback. We also develop a crowdsourcing platform to elicit such signals at a large scale. We explore similar kinds of interactive learning for other agents. In chapter 4 we model the commonly occurring interaction between a driver, the car, and their surroundings. We propose a vehicular sensor-rich platform and learning algorithms to anticipate the future maneuvers of the driver several seconds in advance. In our interaction system, we equip a car with cameras, Global Positioning System (GPS), and a computing device. We observe many drivers driving to their destination, and the drivers provide many implicit examples of maneuvers as learning signals (lane changes, turns etc.). In this dissertation we provide insights into developing such interactive 3 Human Object Object Object Spine Left arm Right arm Left leg Right leg Driver Outsid
ABSTRACT
This study was carried out on the impact of international financial standards (IFRS) on the quality of financia...
Abstract
The study investigates the effects of broken homes on academic performance of secondary school student. The stu...
ABSTRACT
The core northern states represent the birth place of Kannywood, adopting Kano State as the major lead in the formation of the f...
ABSTRACT
Stereospermum kunthianum (Bignoniacea) herb is used in Hausa ethnomedicine in treating bronchitis, veneral diseases, diarrheoa a...
Background Of Study
Since 1981, when the Nigerian Medical Association conducted a meeting in Calabar a...
ABSTRACT
This study attempts to find out the challenges of entrepreneurship in Home and Rural Economics Education. The s...
ABSTRACT
This study determined lead concentrations in seven types of food sold by hawkers in Zaria and environs. Lead is a heavy metal th...
ABSTRACT
This research work has been carried out to determine the challenges of information and Communication Technology...
The study is an analysis of loan administration in financial institution in the nigeria ec...
Abstract
This paper examines the illusion and reality of press freedom in attaining a true democratic system of government in Nigeria. Ar...