PaperlayerView exact revision on arXiv

Source-linked AI summary

Artificial Intelligence and Robotics

Javier Andreu-Perez, Fani Deligianni, Daniele Ravi, Guang-Zhong Yang

arXiv:1803.10813v1cs.AI

TL;DR

The paper examines AI’s evolution, applications, and societal implications, including concerns about privacy, employment, control, and generalisation. It synthesizes historical developments and current capabilities, highlighting major advances while emphasizing that AI remains bounded by data and task specificity. The paper concludes that sustained progress requires responsible attention to ethical and legal challenges.

  • Problem

    AI’s expanding capabilities and applications create unresolved concerns about privacy, employment, loss of control, and generalisation beyond large training datasets.

  • Method

    The white paper provides a comprehensive account of AI’s origins, evolution, current technologies, applications, limitations, and ethical and legal implications.

  • Results

    AI has achieved major advances across applications, including image-recognition error reduction from 28% and 26% to 16% in 2012 and 5% in 2015.

  • Takeaways & Limitations

    AI is expected to become increasingly integrated across sectors, while progress should be sustained through systematic attention to ethical and legal challenges.

  • Takeaways & Limitations

    Current systems often fail in new situations with limited training data because they lack human-like abstraction and generalisability.

Abstract

from arXiv · show

The recent successes of AI have captured the wildest imagination of both the scientific communities and the general public. Robotics and AI amplify human potentials, increase productivity and are moving from simple reasoning towards human-like cognitive abilities. Current AI technologies are used in a set area of applications, ranging from healthcare, manufacturing, transport, energy, to financial services, banking, advertising, management consulting and government agencies. The global AI market is around 260 billion USD in 2016 and it is estimated to exceed 3 trillion by 2024. To understand the impact of AI, it is important to draw lessons from it's past successes and failures and this white paper provides a comprehensive explanation of the evolution of AI, its current status and future directions.

UKRAS.ORG

AI has progressed from specialized reasoning and applications toward more human-like cognitive abilities, while raising questions about its limits and societal impact.

  • AI research spans logic, reasoning, planning, learning, and perception, with intelligence also framed through creativity, emotional knowledge, and self-awareness.
  • Modern AI includes operational and social consequences, with rationality potentially exceeding human performance on specific, well-defined tasks.
  • AI technologies support applications including advertising, driving, aviation, medicine, personal assistance, autonomous vehicles, and game playing.
  • Current AI remains limited to specific applications and lacks common sense, while emotional intelligence is still restricted to detecting basic human states.
  • True and complete AI does not yet exist, although the paper discusses possible human-like cognition and assesses technological, social, ethical, and legal implications.

2. THE BIRTH AND BOOM OF AI

AI developed through foundational advances in computing and neuroscience, followed by cycles of enthusiasm, disappointment, renewed methods, and expanding data-driven capabilities.

  • Early computing and neuroscience connected information processing with the concept of the artificial neuron, establishing foundations for later computational intelligence.
  • 1956 marked the coining of “Artificial Intelligence” at the Dartmouth Conference and the beginning of AI as a named field.
  • 1971–1970s speech-recognition ambitions, funding withdrawals, machine-translation failure, and criticism of perceptrons marked the first AI winter.
  • 1980–1987 expert systems, Japanese fifth-generation computing, and revived connectionism preceded a second winter after hardware-market collapse and maintenance difficulties.
  • AI’s timeline distinguishes positive developments from negative events representing the field’s winters.
  • Since 2000, Big Data, internet and mobile adoption renewed connectionism, while intelligent agents and statistical learning expanded AI applications.

• ALVINN

The section traces AI’s movement from expert systems and specialized task performance toward broader concepts of intelligent agents and strong AI.

  • Expert systems reached an all-time low after the collapse of LISP machines and the end of the US strategic computing program.
  • IBM’s Deep Blue, the first RoboCup, and early robot cars illustrate milestones in specialized machine intelligence and robotics.
  • Weak AI reproduces observed behavior accurately for precision-trained tasks but lacks generalisation, describing most existing machine-learning systems.
  • IBM Watson’s Jeopardy victory is listed among notable milestones in AI’s development.
  • Strong AI would modify its own functioning and perform general intelligent tasks through human-like cognition, but whole-brain simulation remains contested.

3. QUESTIONING THE IMPACT OF AI

The paper examines AI’s potential benefits alongside economic, social, ethical, legal, safety, and governance concerns. It highlights both opportunities for societal services and risks involving autonomy, privacy, warfare, criminal misuse, and human-machine collaboration.

  • The AI100 topics also include public communication, human cognition, neuroscience, machine consciousness, technical forecasting, and delayed translation into socially valuable applications.
  • AI may transform societal services such as health, education, management, and government through economic and social benefits.
  • AI applications raise governance questions involving personal-data privacy, democracy, freedom, law, and ethics.
  • AI’s expanding economic role raises concerns about employment policy and potential financial-market volatility.
  • Autonomous weapons, AI-enabled malware, and terrorist use of drones create security risks requiring boundaries and safeguards.
  • Safe human-robot collaboration requires systems that understand environments and human intentions, not only task precision.
  • Research priorities include formal verification of intelligent-system reasoning, knowledge bases, and behavior within safety boundaries.
  • The paper identifies loss of human control as a major concern requiring technological studies and responsible-development frameworks.

4. A CLOSER LOOK AT THE EVOLUTION OF AI

AI has progressed through alternating periods of optimism and pessimism, with successive advances in expert systems, machine learning, big data, and deep learning. Its international research and patent landscape has shifted toward China and the United States, alongside increased public and private investment and greater industry participation.

  • 4.1 SEASONS OF AI: AI’s history comprises alternating springs of optimism and winters of pessimism from its birth through the present.
  • 4.1 SEASONS OF AI: Machine learning became a cornerstone of AI as the field progressed through specific applications despite limited major funding.
  • 4.1 SEASONS OF AI: Since 2000, internet growth, big data, and deep learning have driven the third spring of AI.
  • Government and public-sector organisations are investing substantially in AI research, including Singapore’s $150 million programme and planned UK funding of £4.7 billion by 2021.
  • Funding for AI safety, ethics, and strategy or policy nearly doubled over three years, with research centres and technology firms investing $6.6 million in 2016.
  • 4.3 PUBLICATION VERSUS PATENTING: China and the United States have led scientific publications and patent filing since 2010, while India and Brazil are rising.
  • Academic AI professionals are increasingly moving into industry, where companies favour preprints and other non-citable documents to avoid peer-review delays.
  • 4.3 PUBLICATION VERSUS PATENTING: Figures 2 and 3 track AI patents and citable articles across countries in five-year periods from 1995 to the present.

5. FINANCIAL IMPACT OF AI

AI’s expanding applications and strong investment have increased its projected economic significance, while major advances demonstrate practical capabilities alongside unresolved limitations.

  • Market growth: $260 billion in 2016 AI-market revenues were projected to exceed $3,060 billion by 2024.The paper links this growth to applications across healthcare, manufacturing, transport, energy, banking, financial services, consulting, government, and advertising.
  • Investment: $1.16 billion was invested in AI start-ups worldwide in 2015, a 10-fold increase since 2009.
  • Technical drivers: Deep learning, improved hardware, sensors, and big data produced significant gains in machine-learning performance and expanded business applications.The paper identifies speech recognition, natural-language processing, and computer vision as major investment areas.
  • Demonstrated progress: AlphaGo defeated professional Go players, marking a historical landmark in AI progress.The paper presents this achievement as evidence that reinforcement-learning systems can exceed human-level performance in some tasks.
  • Limitations and risks: Developing machines with self-sustained long-term goals remains well beyond current technology, amid debate over an AI bubble and insufficient policy frameworks.
  • Demonstrated progress: 5% was the overall error rate of the latest 2015 ImageNet models, compared with 28% and 26% for the contest’s first two years.The paper attributes much of the rapid improvement to GPUs enabling larger models and faster training.

8. HARDWARE FOR AI

AI hardware has shifted from frequency scaling toward multicore and dedicated parallel architectures, with GPUs and FPGAs supporting computationally intensive workloads.

  • CPU evolution: Since 2005, processor progress has emphasized core scaling rather than frequency improvement.Software must be written in a multi-threaded manner to exploit this hardware implementation.
  • Dedicated hardware: Modern AI machines combine multicore CPUs with dedicated parallel hardware, especially GPUs and FPGAs.
  • Dedicated hardware: GPUs accelerate multidimensional-data processing through thousands of smaller cores working independently on input subspaces.
  • Dedicated hardware: FPGAs provide configurable, truly parallel processing by assigning independent tasks to dedicated chip sections.Unlike processors, FPGAs are configured through hardware description languages rather than running memory-stored programs.
  • Historical development: The hardware timeline traces discoveries that influenced AI-system evolution, while CPU history tracks changes in general-purpose computing characteristics.

9. ROBOTICS AND AI

Robotics has progressed from fixed industrial machines toward autonomous, sensor-rich systems that adapt to environments and support perception, interaction, and specialized applications.

  • Evolution and autonomy: Robotics increasingly combines mechatronics, electrical engineering, and computing to produce adaptive sensorimotor functions.
  • Perception and interaction: Computer vision and integrated sensors enable robots to perceive environments, supporting navigation, recognition, planning, and interaction.
  • Perception and interaction: Social robotics covers human–robot interaction and cognitive robotics, including perception of human activities, emotions, and non-verbal communication.
  • Historical milestones: Figure 9 presents a timeline of milestones in robotics and AI.
  • Applications: Robotic applications now include surgical systems, exoskeletons, customer-service robots, drug-delivery nanorobots, and autonomous vehicles.
  • Historical milestones: Milestones include industrial robots, mobile navigation, household robots, humanoids, robot swarms, and systems that defeat humans in games.

10. PROGRAMMING LANGUAGES FOR AI

Programming languages have shaped AI by supporting symbolic manipulation, logic, scientific computation, concurrency, and increasingly deployable machine-learning systems.

  • Foundations: Early AI required languages that could manipulate symbols and lists of symbols rather than only numbers or character strings.
  • Symbolic AI: LISP became influential because of its expressiveness and flexibility, while Prolog was designed to express logical rules and axioms.
  • Scientific and systems programming: C, C++, and Fortran gained popularity in the 1990s for scientific computation, intensive data analysis, and early robotic AI.
  • Modern AI languages: Python became a widely used AI research language because of its versatility and extensive AI, machine-learning, and scientific-computing libraries.
  • Trade-offs: Python remains slower and less efficient than C/C++, Lisp, or Haskell for runtime speed, large-memory management, and highly concurrent systems.
  • Hybrid languages: Hybrid languages such as Scala, Go, Erlang, and Clojure emerged to combine programming paradigms while supporting speed, capacity, concurrency, and parallelization.

11. IMPACT OF MACHINE VISION

Machine vision combines image capture, sensing, and computer-vision algorithms for automated inspection, robot guidance, three-dimensional reconstruction, and object recognition. Deep neural networks and large labelled datasets substantially improved image-classification performance and enabled applications exceeding human accuracy.

  • Machine vision systems: Machine vision integrates image-capture systems with computer-vision algorithms for automatic inspection and robot guidance across 2D and 3D sensing modalities.Sensors include cameras, single-beam lasers, LiDAR, sonar, and other optical systems.
  • 3D reconstruction: Three-dimensional reconstruction commonly uses time-of-flight, multi-view geometry, or photometric stereo to estimate object structure and distance.Time-of-flight measures distance from light-travel time, while multi-view geometry uses corresponding projections and triangulation.
  • Feature extraction: Stereo vision depends on detecting salient features that remain invariant to lighting and robust under geometric transformations.SIFT is described as invariant to scale, rotation, and translation transformations.
  • Object recognition: Object-category recognition is harder to generalise than 3D reconstruction because objects span many categories and may belong to several categories simultaneously.Approaches include deformable parts-and-shape models and bags of visual features represented as words.
  • Deep learning breakthrough: 96% image-classification accuracy in 2015, up from 72% in 2010, followed the emergence of deep neural networks and large labelled datasets such as ImageNet.Deep learning jointly encodes feature extraction and classification, and the reported performance exceeded human accuracy.
  • Collaborative robots: 2017 BWIBots learned human preferences and cooperation by working side by side with people, extending machine vision toward collaborative robotics.The timeline places this development after advances in recognition, sensing, and autonomous systems.

12. ARTIFICIAL INTELLIGENCE AND THE BIG BRAIN

AI development is closely linked to attempts to model brain structure and function, spanning detailed biological simulations, neuromorphic hardware, and neurorobotic systems. These efforts have produced increasingly capable models, while consciousness, intelligence, and biological fidelity remain unresolved questions.

  • Brain-inspired AI: Brain-inspired AI seeks to emulate causal dynamics of internal brain functions so models relate to brain function and behaviour.The approach is debated because the brain is not fully understood.
  • Simulation scales: Neural simulations range from molecular and subcellular processes to individual neurons, local networks, and whole-system models.Computational complexity depends on neuron and synapse counts, network topology, and biological detail.
  • Neural models: Conductance-based Hodgkin-Huxley models capture synaptic interaction and spike generation, whereas integrate-and-fire models trade biological detail for faster simulation.GENESIS and NEURON support systematic modelling of realistic brain networks.
  • Neuromorphic computing: Neuromorphic hardware mimics neuronal structures to achieve functional equivalence and support near-real-time, large-scale neural-network simulation.Neurogrid combines analogue dendritic computation with digital axonal communication.
  • Open questions: The prospect of computers matching the brain’s speed and complexity does not resolve whether simulations would produce consciousness or intelligence.The paper notes that detecting such properties is difficult without thorough definitions of consciousness and intelligence.
  • Brain models and neurorobotics: 2012-present Spaun produced complex behaviour and human performance on simple tasks using a biologically realistic brain model.Spaun ran on the Nengo platform, which specifies collective neural functions alongside electrophysiological details.
  • Research programmes: Brain-inspired projects include cortical-column simulations, artificial-retina systems, closed-loop neurorobotics, and research programmes such as the Human Brain Project.These efforts connect brain simulation, neuromorphic computing, neuroinformatics, and robotic experimentation.

13. ETHICAL AND LEGAL QUESTIONS OF AI

AI and robotics raise ethical and legal questions involving privacy, security, employment, bias, responsibility, and possible civil rights. The paper describes risks from data use, autonomous weaponisation, labour displacement, biased learning, and uncertain liability frameworks.

  • Privacy: AI relies on data as fuel, making privacy protection and safeguards against breaches central responsibilities for AI operators.Applications that compromise privacy may require special legislation.
  • Security and weaponisation: AI-enabled weapons, including autonomous drones, missiles, virtual bots, and malicious software, could contribute to unprecedented war escalation and loss of control.The paper situates these risks within terrorism, regional conflicts, and a global arms race.
  • Employment: 8% of jobs were occupied by robots, with the paper projecting 26% in 2020 as robots gained autonomy and decision-making abilities.The paper links automation risk to possible short-term inequality and the challenge of creating replacement employment.
  • Bias: Machine-learning systems can reproduce human biases, including associations that place flowers near positive terms and insects near negative terms.The paper attributes these associations to links humans made themselves.
  • Bias in applications: Tay produced racist and anti-Semitic remarks after learning from human interactions, while a beauty-contest system eliminated most Black candidates because its training data lacked enough Black people.These examples show that bias can affect both language systems and visual classification.
  • Legal responsibility: Legal proposals include objective civil liability, compulsory insurance, identification of autonomous machines, and responsibility that increases with a robot’s sophistication.The European Parliament’s recommendations address contractual and non-contractual responsibility for robotic systems.
  • Civil rights and labour: Proposals for autonomous robots include electronic-person status, specific rights and obligations, and possible social-security contributions and taxes.These proposals aim to address human–robot relationships and labour-market effects.

14. LIMITATIONS AND OPPORTUNITIES OF AI

AI has achieved strong results in specialised applications, but its current limitations include narrow task scope, weak common sense, difficult generalisation, poor interpretability, and vulnerability to being fooled. Future opportunities include broader sectoral use, brain-inspired systems, and continued investment guided by realistic expectations.

  • Generalisation: AI systems can achieve impressive recognition and speech-translation results after extensive training on large datasets.Their performance often degrades in new situations with limited training data because current systems lack human-like abstraction and generalisability.
  • Interpretability: Deep neural networks often function as black boxes because their millions of parameters make individual decisions difficult to interpret.Visualising high-level features has not generally made trained models interpretable.
  • Robustness: Most current AI systems can be easily fooled, a robustness problem affecting nearly all machine-learning techniques.
  • Opportunities: AI is expected to expand across finance, pharmaceuticals, energy, manufacturing, education, transport, and public services as information and system sophistication increase.The paper describes augmented intelligence as a future stage linking humans and machines.
  • Documented failures: AI failures include robot breakdowns, traffic disruption, false nuclear warnings, mistranslation, fake-news ranking, biased chatbots, and an autonomous-driving fatality.These examples motivate attention to practical limitations alongside AI’s promising results.

15. CONCLUSION AND RECOMMENDATIONS

The paper recommends sustained, responsible development of robotics and AI, combining research investment with attention to social, legal, ethical, economic, and workforce impacts. It calls for coordinated national engagement, stronger research capacity, responsible deployment, and skills development so the UK can benefit from future technological growth.

  • Robotics and AI require a systematic framework that sustains progress while addressing ethical and legal challenges and mitigating adverse effects.The paper argues that fears should motivate responsible development rather than hinder technological progress.
  • Robotics and AI are increasingly important to the UK economy, requiring public engagement about societal change and workforce skill shifts.The recommendation emphasizes a clear and factual public view of current and future developments.
  • A strong research and development base, sustained investment, and support for internationally leading centres are fundamental to UK growth.Funding should prioritize areas of existing strength and projects with greater socioeconomic benefit.
  • Practical deployment and responsible innovation require greater effort to assess economic impacts and maximize benefits while mitigating harms.
  • Government should support workforce adaptation through new opportunities, digital-skills training, re-education, and STEM development for younger generations.These measures are presented as necessary for maintaining UK competitiveness and strengthening robotics and AI expertise.
Loading 1803.10813v1…