Loading field map
Robotics is the branch of artificial intelligence concerned with embodied intelligent action: the design, construction, and control of machines that perceive their environment and act upon it. The field's central question is how a physical system—with sensors, actuators, and limited computational resources—can reliably achieve goals in a world that is partially unknown, dynamic, and uncertain. Robotics sits at the intersection of AI, mechanical engineering, and control theory, but its defining intellectual problem is not mechanical construction; it is the computational problem of deciding what to do next given incomplete information and then executing that decision through noisy hardware.
Every robot, regardless of its form or purpose, instantiates a closed loop: sense, model, plan, act. Sensors—cameras, laser rangefinders, inertial measurement units, tactile arrays, and many others—provide partial and noisy measurements of the world. Actuators—wheels, legs, arms, grippers—apply forces that change both the robot's state and the state of the world around it. Between sensation and action lies the robot's internal reasoning, which may range from a simple reflex to an elaborate hierarchical planner. The central difficulty is that the loop is never clean: sensors have noise and limited range, actuators slip and drift, the world changes while the robot computes, and the robot's own actions perturb the very environment it is trying to measure.
Modern robotics treats this problem probabilistically. Almost no quantity a robot needs—its own position, the location of an object, the map of a building—is known with certainty. The dominant computational framework, probabilistic robotics, maintains an explicit belief over the robot's state and the world's state, updates that belief as new sensor data arrives, and chooses actions that reduce uncertainty while making progress toward the task. This framework gives robotics a distinctive character within AI: where much of AI research concerns reasoning about abstract information in a stable digital world, robotics must contend with the fact that the world is not available as a database, only as a stream of noisy, partial observations.
From the 1960s through the 1980s, robotics research was dominated by the sense–model–plan–act paradigm, sometimes called the deliberative or classical approach. A robot first built a complete internal model of its environment, typically a geometric map or a logical description, then derived a course of action through a general-purpose planner, and finally executed the plan by translating it into motor commands. This approach inherited its conceptual structure from early AI's emphasis on symbolic reasoning: perception was treated as a translation problem (sensor data into symbols), planning as a search problem over states, and execution as a bookkeeping problem.
The classical paradigm achieved notable success in controlled settings, most famously in industrial manufacturing, where a robot arm could be programmed with a precise geometric model of its workspace and could repeat the same motion indefinitely. But it failed conspicuously in unstructured environments. The assumptions that made planning tractable—a complete model, predictable dynamics, static surroundings—did not hold outside the factory. A robot could spend minutes building a map that was already obsolete, or produce a plan that failed because a door had drifted a few degrees. These failures were instructive: the bottleneck was not computation but representation. The world is too rich and too variable to be captured in advance, and the time spent building a complete model is time not spent acting.
In the mid-1980s, a reactive movement, now usually called behavior-based robotics, challenged the classical program directly. Its core claim was that intelligence need not require a central model. Instead of a single planner that consulted a unified world model, a robot could run many simple behaviors—avoid obstacle, approach goal, track target—in parallel, with arbitration mechanisms deciding which behavior controlled the actuators at any moment. This architecture, often associated with Rodney Brooks's subsumption approach, built robots that navigated offices and cluttered rooms using little more than direct sensor–actuator mappings.
The behavior-based school's contribution was twofold and durable. First, it demonstrated that surprisingly competent behavior could emerge from simple, decentralized rules, without any explicit representation of the environment. Second, it reframed the central question of robotics: rather than asking how to build the most accurate model, it asked how to couple perception tightly to action so that a robot could respond quickly to what actually mattered. Its limits were equally instructive. Behavior-based robots excelled at tasks that were inherently local and short-horizon—moving through a room, following a wall—but struggled with tasks requiring global consistency, such as navigating across an entire building or manipulating tools with multiple stages. A robot with no map cannot know whether it has already visited a corridor, and a robot with no representation of an object cannot reason about how to use it to reach a goal.
Today the deliberative and reactive approaches are best understood not as competing camps but as opposite ends of a spectrum. Nearly all modern robots use layered architectures that combine reactive behaviors for safety and fast response with deliberative reasoning for global coordination. The debate of the 1980s settled into a pragmatic synthesis: the question is no longer whether a robot should have a model, but what granularity of model, over what horizon, and with what mechanism for keeping it in touch with reality.
From the 1990s onward, robotics absorbed a third intellectual current that reconciled much of the earlier conflict. Probabilistic robotics treats the robot's knowledge not as a complete model or as an absence of models, but as a probability distribution over possible worlds. The robot's belief—its estimate of where it might be, what the map might look like, where an object might lie—is represented explicitly and updated with Bayes' rule as measurements arrive. The result is that uncertainty is not an error to be eliminated but a quantity to be managed.
This framework gave rise to the two success stories that define the modern field's applied core. The first is simultaneous localization and mapping (SLAM), the problem of building a map of an unknown environment while simultaneously determining the robot's position within it. SLAM is a circular problem—a map is needed to localize, and a pose is needed to build a map—but probabilistic methods resolve it by treating both jointly as a single estimation problem. Over the 2000s, SLAM matured from a research problem into a deployed capability: most self-driving cars, drones, and mobile robots in warehouses use some variant of it. The second is motion planning under uncertainty, where a robot plans a trajectory not over assumed states but over its belief distribution, choosing actions that minimize expected cost while reducing the uncertainty that threatens task success.
Probabilistic robotics did not abolish the tension between reactivity and deliberation; it gave the field a common language in which to discuss it. A reactive behavior is, from this perspective, a policy that maps beliefs directly to actions without deliberative search, justified either by the cost of deliberation or by the fact that the task requires too little lookahead to merit it. A deliberative planner is a policy that spends computation on search because the payoff of better global decisions outweighs the cost. The choice is now an engineering trade-off, addressed with formal tools, rather than a philosophical commitment.
The most recent major shift in robotics is the integration of machine learning, particularly deep learning and reinforcement learning, into nearly every stage of the perception–action loop. Classical approaches required engineers to hand-design feature detectors for vision, state estimators for dynamics, and cost functions for planning. Learning-based approaches instead train neural networks on large datasets—or on experience gathered by the robot itself—to perform these functions. A deep network can learn to detect objects in raw camera images, to estimate the pose of a robot's hand from visual input, or to map sensor readings directly to motor commands.
This shift has been most visible in manipulation, the subfield concerned with hands, grippers, and the physical interaction with objects. For decades, manipulation was dominated by analytic grasp planning: given a geometric model of an object and a hand, compute a contact configuration that will hold the object stably. This approach struggled with the enormous variety of real objects and the difficulty of modeling friction and compliance. Learning-based manipulation instead collects thousands of grasping attempts—in simulation or on real robots—and trains a policy that predicts whether a proposed grasp will succeed, often from raw vision without any model of the object.
The limits of the learning-based approach are as important as its successes. Learned policies are only as good as their training data, and robotic data is expensive to collect. A simulation-trained policy may fail in the real world because the discrepancy between simulation and reality is large. A policy trained in one kitchen may not transfer to another with different lighting, colors, or object placements. The field's current frontier is precisely the problem of generalization: how to train policies that are robust to the variation of the physical world, and how to acquire enough data to train them. This has led to intense interest in simulation as a training ground, in domain randomization (training over many randomized simulations to induce robustness), and in using large, heterogeneous datasets gathered across many robots and tasks.
Reinforcement learning—where a robot learns by trial and error, receiving reward signals for successful behavior—has produced remarkable isolated demonstrations, such as robot hands learning to manipulate objects or legged robots learning to walk. Its general deployment remains constrained by sample efficiency (real robots simply cannot attempt millions of trials) and by safety (an exploring robot may damage itself or its environment). A common practical compromise is a hybrid: use model-based reasoning or a classical controller for structure and safety, and use learned components for perception or for the parts of the policy that are hard to derive analytically.
The modern landscape is thus best understood as a division of labor. Classical planning provides guarantees and interpretability over long horizons. Probabilistic estimation provides a principled treatment of uncertainty. Reactive behaviors provide robustness to unmodeled detail. Learning provides the ability to extract structure from data that is too complex to hand-program. The field no longer has a single dominant doctrine; its research frontiers concern how to combine these tools, how to make learned components reliable, and how to bridge the gap between the controlled conditions in which methods are developed and the uncontrolled conditions of the real world.
Beneath the shifting methods, robotics is organized by a small set of enduring questions that any robot, of any era, must answer. The first is the perception problem: given raw sensor data, how much of the world's structure can be recovered, and at what cost? The second is the representation problem: what should a robot remember about the world, in what form, and when is memory not needed? The third is the planning problem: how far ahead must a robot reason, and over what space of actions? The fourth is the execution problem: how can commands be translated into forces and torques that achieve the intended effect despite friction, slip, and unexpected contact? The fifth is the learning problem: what can be acquired from experience, rather than specified in advance? Different eras and schools have given different weights to these questions, but no approach has eliminated any of them, because each corresponds to an inescapable property of the physical world: the world is not given, the world is not static, the world is not fully known, and the robot acts through imperfect hardware.