Loading field map
Mobile robotics is the branch of robotics concerned with machines that can move through and operate within their environment, as opposed to fixed manipulators anchored to a single base. The central problem of the field is not simply making something move, but making movement purposeful: a mobile robot must decide where to go, figure out how to get there, and execute that plan while coping with an imperfectly known, changing world. This requires integrating mechanics, sensing, computation, and control into a single system that can act with some degree of autonomy.
The defining challenge of mobile robotics is that the robot's knowledge of its own state and surroundings is always incomplete and noisy. A wheeled robot may think it has traveled ten meters, but wheel slip makes that estimate unreliable. A camera may show a corridor, but the robot cannot know what lies around the next corner. A laser scanner may detect a wall, but cannot tell whether the wall is load-bearing or a flimsy partition. The field is therefore organized around a set of interconnected questions: How can a robot determine where it is? How can it build or update a model of its environment from sensor data? How can it plan a path through that environment? And how can it execute that plan robustly when the world does not match the model?
These questions are not solved once and then forgotten. They are continuously re-solved as the robot moves, because every action changes the robot's state and every new sensor reading changes its beliefs. The practical consequence is that mobile robots are rarely built as a single monolithic program. Instead, they are typically structured as a cycle of perception, state estimation, planning, and control, with each stage feeding into the next. The field's history is largely the story of how researchers learned to make each stage more robust and how they learned to integrate the stages without letting errors accumulate catastrophically.
The most fundamental question in mobile robotics is also the most deceptively simple: Where am I? A robot that cannot answer this question cannot meaningfully plan or act. The difficulty is that the answer depends on the robot's frame of reference. In a known environment, a robot can use landmarks—distinctive features it can detect with its sensors—to triangulate its position. But in an unknown environment, the robot must simultaneously figure out where it is and what the environment looks like. This is the simultaneous localization and mapping (SLAM) problem, and it is the intellectual core of the field.
SLAM is a chicken-and-egg problem: a map is needed to localize, but localization is needed to build a map. The breakthrough insight, developed in the 1980s and 1990s, was that the two problems must be solved jointly, with the robot maintaining a probability distribution over both its own pose and the positions of landmarks. The standard mathematical framework treats the robot's state and the map as a single high-dimensional random variable, updated with each motion and each sensor measurement. Early solutions used extended Kalman filters, which linearize the problem around the current estimate; later approaches used particle filters, which represent the probability distribution with a set of weighted samples, or graph-based methods, which treat the problem as one of optimizing a network of constraints.
The practical significance of SLAM cannot be overstated. It is the reason a robot vacuum can build a floor plan of a home it has never seen, the reason an autonomous vehicle can navigate a city street without a preloaded map, and the reason a drone can explore an indoor space where GPS is unavailable. But SLAM is not a single solved problem. It has many variants—visual SLAM using cameras, lidar SLAM using laser scanners, and hybrid approaches—and each has different strengths and failure modes. Visual SLAM is cheap and information-rich but struggles in low light or featureless environments; lidar SLAM is robust and accurate but expensive and limited in range. The choice of sensor and algorithm depends on the application, the environment, and the computational resources available.
Once a robot has a map and a position, it must decide how to reach a goal. Path planning is the problem of finding a route from the current location to a target location, typically while avoiding obstacles and satisfying constraints such as minimum energy or maximum safety. The classical approach is graph-based search: the environment is discretized into a grid or a set of waypoints, and an algorithm such as A* searches for the shortest path through the graph. This works well in static, known environments, but real environments are rarely static. People move, doors close, and obstacles appear.
The response to this limitation has been a family of techniques that plan in a way that is robust to uncertainty. The rapidly exploring random tree (RRT) family of algorithms grows a tree of feasible paths by sampling random points in the environment, making them well suited to high-dimensional spaces and nonholonomic vehicles that cannot turn in place. Probabilistic roadmaps precompute a network of feasible paths that can be reused for multiple queries. More recent approaches incorporate the robot's uncertainty directly into the planning objective, choosing paths that minimize the expected cost of collision or that keep the robot in areas where its localization is likely to remain accurate.
Planning is not a one-time computation. A mobile robot must replan as it moves, because new sensor data may reveal that the planned path is blocked or that a better path exists. This has led to the notion of the perception–planning–control loop as a continuous process rather than a sequence of discrete stages. The robot is always sensing, always updating its beliefs, and always revising its plan in response.
A plan is only as good as the robot's ability to execute it. The control problem is to generate commands to the robot's actuators—wheel motors, thrusters, legs—that cause it to follow the planned trajectory. For wheeled robots, this is complicated by nonholonomic constraints: a car cannot move sideways, and a differential-drive robot cannot move directly toward a point without first turning toward it. Control theory provides a rich set of tools for this problem, from simple proportional–integral–derivative (PID) controllers that correct errors between desired and actual state, to model predictive control, which solves a finite-horizon optimization problem at each time step to find the best sequence of commands.
The deeper challenge is that the robot's model of its own dynamics is never perfect. Friction, tire slip, battery voltage, and surface conditions all affect how the robot actually moves. Robust control methods explicitly account for these uncertainties, designing controllers that are guaranteed to work within a bounded range of model error. Learning-based control methods, by contrast, attempt to learn the robot's dynamics from data, either offline or online, and use that learned model to improve performance.
A major historical divide in mobile robotics concerned the proper role of internal models and planning. The deliberative approach, dominant in the early years of the field, held that a robot should build a detailed model of its environment, plan a course of action, and then execute that plan. This approach worked well in controlled settings but proved brittle in the real world, where the model quickly becomes stale and the planning computation can be too slow to respond to immediate dangers.
In the 1980s, a reaction against this view emerged, most prominently associated with Rodney Brooks and the behavior-based approach. Brooks argued that intelligence does not require representation, and that a robot could achieve robust behavior by layering simple reactive behaviors—avoid obstacles, wander, seek light—that operate directly on sensor data without any central model or planner. This approach, known as subsumption architecture, produced robots that moved surprisingly well in cluttered environments, and it sparked a vigorous debate about the role of representation in robotics.
The debate was never fully resolved, and in retrospect it is better understood as a productive tension than a victory of one side over the other. Modern mobile robots almost always combine reactive and deliberative elements. A robot vacuum uses a reactive bump sensor to avoid immediate collisions but also maintains a map and plans coverage paths. An autonomous vehicle uses reactive emergency braking but also plans long routes and predicts the behavior of other road users. The field has largely converged on a pragmatic hybrid: reactive layers for safety and immediacy, deliberative layers for goal-directed behavior, and a middle layer that mediates between them.
A more recent shift has been the incorporation of machine learning into nearly every stage of the mobile robotics pipeline. Learned perception systems, particularly deep neural networks, have dramatically improved the ability of robots to interpret raw sensor data—recognizing objects, detecting drivable surfaces, estimating depth from a single camera. Learned control policies, trained through reinforcement learning, have produced robots that can walk, run, and manipulate objects with a dexterity that would be extremely difficult to program by hand.
The appeal of learning is that it offers a way around the fundamental difficulty of modeling the world. Instead of writing rules for every possible situation, the robot learns from data what to do. The cost is that learned systems are opaque, difficult to verify, and can fail in unexpected ways. A neural network that drives well in sunny conditions may fail catastrophically in fog, and it is often unclear why. This has led to a growing emphasis on safety and verification in learning-based robotics, and to hybrid approaches that combine learned perception with classical planning and control.
The relationship between learning and classical methods is not one of replacement. Learned systems are often used to improve components of a classical pipeline—a learned object detector feeding a classical planner, for example—rather than to replace the pipeline entirely. And classical methods remain essential for the parts of the problem that require guarantees, such as collision avoidance and stability. The current landscape is one of integration, with the choice of method driven by the requirements of the task, the availability of data, and the need for interpretability and safety.
Mobile robotics today is a mature field with a well-established set of core techniques and a rapidly expanding set of applications. The fundamental problems—localization, mapping, planning, control—remain the organizing structure of the field, and the standard solution methods are taught in every robotics curriculum. But the field is also being reshaped by advances in sensing, computation, and learning, and by the demands of real-world deployment.
The most visible applications are in transportation and logistics. Autonomous vehicles, delivery robots, warehouse robots, and agricultural machines all rely on the core techniques of mobile robotics, adapted to their specific environments and requirements. The challenges of these applications are not just technical but also social and regulatory: how to ensure safety, how to handle edge cases, how to interact with humans. These concerns have pushed the field toward greater attention to robustness, verification, and human–robot interaction.
Underlying all of this is a set of enduring questions that the field has not fully answered. How can a robot know what it does not know? How can it act safely when its model of the world is wrong? How can it learn from experience without losing the guarantees that classical methods provide? These questions are unlikely to be settled once and for all. They are the permanent frontier of a field whose defining feature is the attempt to build machines that can cope with the open-ended complexity of the real world.