Loading field map
A mobile robot that moves through an unconstrained environment must solve a deceptively hard problem: it has to decide where to go and how to get there while the world is cluttered, dynamic, and only partially observable through noisy sensors. The history of mobile robotics can be understood as a sequence of competing answers to this challenge, each framework exposing the blind spots of its predecessors and carving out its own territory. Seven major frameworks have emerged, and most remain active today, having settled into a pragmatic division of labor.
The earliest approach to mobile robot navigation, Deliberative Mobile Navigation (1966–Present), treated the problem as a chain of sense–plan–act. The robot builds a complete map of its environment, plans a path to the goal using that map, and then executes the plan. The landmark Shakey robot at SRI International embodied this model: it used cameras and range sensors to build a symbolic world model, then reasoned logically to generate a sequence of actions. Deliberative navigation promised optimality and explainability, but it collapsed in practice because the real world rarely matched the model. Sensor noise, moving obstacles, and imperfect actuation meant that plans quickly became obsolete. The framework’s reliance on a perfect global model made it brittle outside tightly controlled settings.
By the mid-1980s, researchers had grown frustrated with the fragility of deliberate planning. Reactive Obstacle Avoidance (1986–Present) abandoned the idea of building a world model altogether. Instead, it used local sensor readings to generate immediate motor commands: if a sonar detects an obstacle to the left, turn right. Techniques such as the Vector Field Histogram (VFH) and potential fields allowed robots to dart through cluttered spaces at high speed without any internal representation. The trade-off was stark—reactive systems were fast and robust to unexpected obstacles, but they had no memory of where they had been and no way to plan toward a distant goal. They could avoid a chair but not navigate from a kitchen to a bedroom.
At almost the same moment, Behavior-Based Mobile Robotics (1986–Present) offered a different way to escape the deliberative dead end. Instead of a single sense–plan–act loop, behavior-based systems decomposed navigation into parallel, task-specific behaviors—wander, avoid, follow wall, go to goal—that compete or cooperate to produce action. The subsumption architecture, introduced by Rodney Brooks, was the most radical expression: it built behavior layers that inhibited lower layers, creating a tight coupling between sensing and action. Where reactive obstacle avoidance was a technique, behavior-based robotics was a philosophical commitment: intelligence emerges from the interaction of simple, situated behaviors, not from an internal world model. Both frameworks rejected the representational core of deliberative navigation, but behavior-based systems could be extended to longer-term goals by adding behaviors that remembered past states, whereas pure reactive systems could not.
The obvious next step was to combine the long-horizon reasoning of deliberative systems with the fast reaction of reactive ones. Hybrid Deliberative-Reactive Navigation (1987–Present) layered a deliberative planner on top of a reactive executor. The planner generated a coarse route—typically a sequence of intermediate goals or a topological path—while the reactive layer handled local obstacle avoidance and moment-to-moment control. Architectures such as AuRA (Autonomous Robot Architecture) explicitly separated the two levels, with the planner updating only when the robot sensed it had deviated too far from the intended route. Hybrid systems did not simply “combine” the earlier approaches; they redefined them. Deliberative planning was scaled down to a loose sketch, not a precise trajectory, and reactive control was embedded within that sketch. The result was a pragmatic compromise that dominated fielded robots for decades.
While hybrids addressed the speed–planning trade-off, they still assumed that the robot’s sensors gave reliable information about the world. That assumption was false. Probabilistic Localization and Mapping (1985–Present) recognized that uncertainty is inescapable and should be modeled explicitly. Instead of maintaining a single best estimate of the robot’s position, probabilistic methods represent a probability distribution over possible poses. Monte Carlo Localization (MCL), introduced in the late 1990s, uses a particle filter to approximate this distribution, allowing robots to localize even when they become lost. Simultaneous Localization and Mapping (SLAM) extended the idea: the robot builds a map of the environment while simultaneously keeping track of its own location within that map, all while treating both as uncertain quantities. Probabilistic methods did not replace hybrid architectures; they became the infrastructure that made them reliable. A modern hybrid system typically uses a probabilistic localizer to keep the planner informed of the robot’s true state, and the reactive layer is often probabilistic as well.
Early mobile robots relied on wheel odometry and sonar or laser range finders for perception. Visual Navigation (1987–Present) shifted the primary sensor to cameras, which are cheap, passive, and rich in information. The challenge is that images are ambiguous and high-dimensional. Early visual navigation followed the same dichotomy as earlier work: map-based approaches built metric or topological maps from visual landmarks, while mapless approaches used qualitative visual features to steer toward a goal. The real breakthrough came when visual navigation converged with probabilistic methods. Visual SLAM uses camera images to perform the same uncertainty-aware mapping and localization that laser-based SLAM had achieved, but with the added difficulty of extracting stable features from video streams. This marriage challenged the assumption that metric maps were necessary: visual SLAM often produces sparse feature maps or topological graphs that are sufficient for navigation. Today, visual navigation is the dominant sensing modality for consumer drones, autonomous cars, and many mobile robots.
The most recent framework, Learning-Based Mobile Robotics (1995–Present), challenges a core assumption shared by all earlier approaches: that the robot’s navigation logic must be hand-designed. Instead, learning-based methods use data—either from human demonstration, self-exploration, or simulation—to train navigation policies. Early work, such as Sebastian Thrun’s 1995 use of neural networks and reinforcement learning to drive a robot through office corridors, showed that a robot could acquire navigation skills directly from experience. The learning-based approach does not reject the insights of earlier frameworks; it often absorbs them as components. A learning system might use a probabilistic localizer for state estimation and a hybrid planner for goal reasoning, but it learns the low-level control policy or the cost function that makes the planner work in novel environments. The most dramatic shift is in the treatment of models: where earlier frameworks assumed that the robot’s dynamics and the environment’s structure were known (or could be estimated), learning-based methods treat these as unknowns to be discovered from data.
All seven frameworks remain active in contemporary mobile robotics, but they have settled into a pragmatic division of labor. Deliberative planning survives in the high-level route planning layer of autonomous vehicles and warehouse robots. Reactive obstacle avoidance is the default for collision avoidance in every fielded system. Behavior-based architectures are less common as standalone frameworks but the ideas of behavior composition and arbitration persist in many robot control stacks. Hybrid deliberative-reactive navigation is the de facto standard architecture for most real-world robots. Probabilistic localization and mapping is a core component in virtually every robot that moves autonomously. Visual navigation has largely replaced laser-based sensing for outdoor and low-cost platforms. Learning-based methods are growing rapidly, especially in perception (learning to detect obstacles) and in end-to-end control for specific tasks.
Where the frameworks agree is that uncertainty must be managed, that no single layer can solve navigation alone, and that the robot’s behavior must be robust to a world that does not match its expectations. Where they disagree is on how much modeling versus learning is appropriate, whether metric maps or topological representations should be used, and whether deliberative reasoning needs to be explicit or can be learned. The leading frameworks today—Probabilistic Localization and Mapping, Hybrid Deliberative-Reactive Navigation, and Learning-Based Mobile Robotics—are not competitors but collaborators. The typical modern system uses a probabilistic core for state estimation, a hybrid architecture for decision-making, and learning for perception, control, or both. The next chapter will likely be about how learning reshapes the division of labor further, perhaps absorbing parts of the deliberative layer entirely.