Artificial Intelligence, Sub-fields, and Embodied Robotics

Artificial intelligence is the study and engineering of systems that perceive their environment and choose actions to achieve goals, including learning, reasoning, planning, and language use, which we normally associate with intelligent behavior in humans or animals.

AI has no single agreed definition. The standard textbook, Russell and Norvig's Artificial Intelligence: A Modern Approach, organizes the many definitions into four approaches along two axes: whether a system is judged by its thinking or its behavior, and whether the benchmark is human performance or an ideal (rational) standard:

• Thinking humanly (cognitive modeling): build programs that reproduce how human minds work.
• Thinking rationally (the "laws of thought"): formal logic and correct inference.
• Acting humanly: the Turing Test (Turing, 1950).
• Acting rationally: the rational agent, which does the best expected thing given what it knows.

Modern AI mostly follows the rational agent view. It is more general than the laws-of-thought approach, because correct inference is only one way to act rationally. It is also easier to define and measure scientifically than human thought or behavior. An agent is simply something that receives percepts and performs actions, and intelligence is judged by how well those actions achieve the agent's objectives.

• Machine learning (ML): Systems improve at a task from data rather than from hand-written rules.
• Natural language processing (NLP): Computers understand, generate, and translate human language. Examples: machine translation, chatbots, summarization, speech recognition. Large language models belong here. This is especially relevant for multilingual settings, where handling code-switching and low-resource languages such as Arabic dialects remains hard.
• Robotics: AI is embodied in machines that sense and act in the physical world. It combines perception, planning, control, and learning.
• Knowledge representation and reasoning (KRR): Encoding facts and rules so systems can infer new knowledge: logic, ontologies, knowledge graphs, expert systems.
• Planning and search: Finding sequences of actions that reach goals, such as route planning, scheduling, and game-tree search.
• Multi-agent systems and AI ethics/safety: These cover coordination between agents, fairness, transparency, and alignment of AI systems with human values.

The field's core ideas are old. Neural networks date to the 1940s–60s, and backpropagation was popularized in the 1980s. What changed is that several bottlenecks were removed at about the same time:

• a) Data: Learning-based methods are data-hungry. The internet, smartphones, sensors, and social media produced enormous datasets. Large labeled datasets such as ImageNet (millions of labeled images) gave deep learning something to learn from.
• b) Compute: GPUs, originally built for graphics, turned out to be well suited to the matrix arithmetic of neural networks. Specialized hardware (TPUs and AI accelerators) and cloud computing then made large-scale training available to many more groups.
• c) Algorithms and architectures: The Transformer architecture (Vaswani et al., 2017), which scales well and underlies modern large language models, alongside empirical scaling laws (Kaplan et al., 2020), showing that performance improves predictably with more data, parameters, and compute, which encouraged very large investments.
• d) Economic incentives and accessibility: Companies found profitable applications (search, recommendation, ads, translation), so funding grew. Public interfaces released in the early 2020s made AI visible to non-experts, which accelerated adoption and attention.

Earlier waves hit limits: the 1970s and the late-1980s/early-1990s "AI winters" followed over-promising, brittle rule-based expert systems, and insufficient computing power and data.

• Challenge 1: Robust localization and navigation in a large, dynamic, crowded environment
  • GPS is unreliable indoors: Research on security robots notes that satellite navigation works outdoors but is blocked indoors by building structure, so the robot must rely on onboard sensing such as LiDAR, cameras, wheel odometry, and IMUs. Even outdoors near buildings, GPS accuracy degrades because the buildings occlude satellites (Capezio et al., 2007, localization for an airport night-surveillance robot).
  • Dynamic obstacles: Crowds, strollers, children, and cleaning carts require real-time obstacle avoidance and social navigation, meaning predicting pedestrian motion and passing politely.
  • Wheeled-robot limits: Stairs, escalators, thresholds, and glass walls (which confuse LiDAR and cameras) restrict where the robot can go. Large open atria and long, repetitive corridors also cause perceptual aliasing, where different places look alike and the robot "gets lost."
  • Failure modes: Localization drift accumulates over hours, so the system needs loop closure, relocalization after being kidnapped (picked up or blocked), and graceful behavior when lost.
  • Typical technical responses: SLAM (simultaneous localization and mapping), sensor fusion with probabilistic filters, continuous map updating, semantic maps, and fallback to remote operator help (Thrun et al., 2005).
• Challenge 2: Energy, endurance, and reliability over 24 hours
  • Battery limits: A single charge rarely lasts a full day, so the robot must autonomously return to a charging dock, ideally at low-traffic times, without leaving coverage gaps. Some surveillance platforms address this with intelligent battery charging systems, and fleet designs use several robots that hand over patrol zones.
  • Mechanical wear and maintenance: Continuous operation stresses wheels, motors, and sensors. Dust, spills, and polished floors affect traction and sensors, so the robot needs self-diagnosis and predictive maintenance.
  • Connectivity and software faults: Wi-Fi dead zones, network outages, and software crashes need watchdogs and safe-stop behavior. A robot that freezes in a main corridor is itself a safety and reputation problem.
• Challenge 3: Reliable perception, decision-making, and human acceptance (security intelligence)
  • Detecting what matters: The robot must separate normal activity from suspicious events (abandoned bags, restricted-area entry, fights, fires, spills, after-hours intruders) across daytime crowds and night-time empty halls. Lighting varies from bright atria to dark corridors.
  • Human-robot interaction and safety: The robot must be safe around the public: low speed, collision-avoidance redundancy, an emergency stop, and a clear explanation of what it is doing. People may also harass, block, or vandalize it.
  • Privacy, ethics, and law: Continuous video recording of shoppers raises data protection concerns: lawful basis, notices, retention limits, access control, and limits on facial recognition. The robot's value depends on public trust.