Mobile Robotics

Mobile robotics is the branch of robotics concerned with machines that move through the world on their own: wheeled rovers, walking robots, drones, underwater vehicles, and the software that lets them find their way.

7 frameworks, from 1940 to today · Robotics · Engineering

still a working view ended · each leads to its entry below

Overview

Mobile robotics is the branch of robotics concerned with machines that move through the world on their own: wheeled rovers, walking robots, drones, underwater vehicles, and the software that lets them find their way1. What makes it distinct from other robotics is that the robot's environment is not fixed. A factory arm works in a cell whose geometry an engineer has measured to the millimetre; a mobile robot must cope with a world it has never fully seen, with people who step into its path, with floors that change, and with its own imperfect knowledge of where it is. The field's history is a sequence of answers to one question: how much of the world should the robot be told about in advance, and how much should it work out for itself as it goes?

Teleoperation and the first autonomous rovers, 1940s–1960s

The earliest mobile robots were not autonomous at all. During and after the Second World War, remote-controlled vehicles and manipulators were developed for handling hazardous material, and the pattern of a human operator driving a machine at a distance persisted into the 1950s and 1960s. The first steps toward autonomy came from a different direction: the desire to explore places people could not easily go.

In the late 1960s, the Stanford Research Institute built Shakey, often described as the first mobile robot to reason about its actions12. Shakey was a box on wheels with a television camera, a rangefinder, and a bump sensor, and it was controlled by a program that decomposed a task into subgoals, planned a route, and executed it step by step. Shakey moved slowly and needed a carefully prepared environment, but it established the template that would dominate the field for two decades: sense, plan, act. The robot would perceive the world, build an internal model of it, compute a plan against that model, and then carry the plan out.

The deliberative paradigm and its limits, 1970s–1980s

The sense–plan–act architecture, sometimes called deliberative robotics, treats navigation as a problem of reasoning. The robot maintains a map, locates itself on the map, and searches for a path to its goal. This approach was given mathematical form in the 1970s and 1980s by work on configuration space—the space of all possible positions and orientations of the robot—and by path-planning algorithms that searched that space for collision-free routes. The appeal was that a plan, once computed, could be checked for safety before the robot moved.

Deliberation worked well in structured settings, and it produced some of the field's landmark demonstrations. But it had a structural weakness: planning is slow, and the world does not wait. A robot that stops to recompute a route whenever a person walks past will spend most of its time thinking rather than moving. Worse, the map the robot plans against is always out of date, because the world changes faster than the robot can update its model. By the late 1980s, a generation of researchers had concluded that building a complete world model and reasoning over it was the wrong foundation for a machine that had to survive in real time.

Reactive control: potential fields and behaviour-based robotics, 1980s–1990s

Two responses to the deliberative bottleneck appeared almost simultaneously, and both shared the idea that the robot should react to its sensors directly rather than through a central model.

Artificial potential fields, developed by Oussama Khatib in the mid-1980s, treated the robot as a particle moving under forces. The goal attracted the robot; obstacles repelled it; the robot simply followed the resulting force vector. The method was elegant and fast—no search was needed—and it could be computed continuously as the robot moved. Its weakness was equally clear: a robot can get trapped in a local minimum, a spot where the attractive and repulsive forces cancel even though the goal lies elsewhere, such as the far side of a U-shaped obstacle. Potential fields remained a standard tool, but they were understood as a local method that needed something else to escape traps.

Behaviour-based robotics, associated above all with Rodney Brooks at MIT in the mid-1980s, went further3. Brooks argued that intelligence did not require a central representation of the world at all. Instead, a robot could be built from a set of simple behaviours—avoid obstacles, follow a wall, move toward a goal—each wired directly from sensors to actuators, with a priority scheme deciding which behaviour took control at any moment. His robot Genghis, a six-legged walker built in the late 1980s, walked over rough ground without any map or planner, using reflexes coordinated at the level of individual legs. The approach was influential far beyond navigation: it changed how roboticists thought about the relationship between sensing, acting, and representation, and it seeded work in multi-robot coordination, where simple local rules among many robots produced group behaviour no single robot planned.

The debate between deliberative and reactive camps was genuine and lasted through the 1990s. Deliberators pointed out that reactive robots could not plan a route through a large building or explain why they had done what they did. Reactivists pointed out that deliberative robots were too slow and too fragile for the real world. The resolution, when it came, was to combine them.

Probabilistic robotics: navigating under uncertainty, 1990s–2000s

The synthesis rested on a change in what the robot's internal model was for. Instead of a single best-guess map, the robot would maintain a probability distribution—a set of possible states of the world, each with a likelihood attached. This is the core of probabilistic robotics, which matured through the 1990s and was codified in the textbook of that name by Sebastian Thrun, Wolfram Burgard, and Dieter Fox in 2005.

The key insight was that a mobile robot is never certain where it is. Wheel odometry drifts; sensors are noisy; the world contains repeated features such as identical corridors. Probabilistic methods handle this by updating beliefs as evidence arrives. The particle filter, a technique in which a cloud of sampled hypotheses is weighted by how well each matches the sensor readings and then resampled, became the standard tool for localization—the problem of determining the robot's position on a known map. The same machinery, run in reverse, allowed a robot to build a map while simultaneously localizing itself on it, a problem known as SLAM (simultaneous localization and mapping)4. SLAM was one of the defining research programmes of the 1990s and 2000s: early solutions used extended Kalman filters, and later particle-based and graph-based methods made it practical on ordinary computers. Probabilistic techniques also transformed planning, allowing robots to reason about the likelihood of collision and about the cost of being wrong.

The practical consequence was that robots could at last operate in buildings and streets that had not been specially prepared for them. Probabilistic navigation became the backbone of the autonomous vehicles and service robots that followed.

Field and challenge-driven autonomy, 2000s–2010s

The decisive test of these methods came from competitions. The U.S. Defense Advanced Research Projects Agency (DARPA) ran a series of Grand Challenges for autonomous ground vehicles, beginning in 20042. The first event ended with no vehicle completing the desert course; by 2005, five vehicles finished, and the winning entry from Stanford, led by Sebastian Thrun, combined probabilistic mapping, machine learning for terrain classification, and a planning layer that chose among candidate trajectories. The Urban Challenge of 2007 moved the problem into a mock city, requiring vehicles to obey traffic rules and interact with other traffic. These events forced the field to integrate perception, mapping, planning, and control into complete systems, and they demonstrated that the combination worked.

The same period saw mobile robotics move into the field at scale: planetary rovers such as NASA's Spirit and Opportunity, which landed in 2004 and drove for years using onboard navigation and hazard avoidance; agricultural robots; warehouse robots; and the first generation of autonomous road vehicles. The pattern that emerged was a layered architecture—a probabilistic map and localizer underneath, a planner above, and a reactive safety layer that could override both—which reconciled the deliberative and reactive traditions rather than choosing between them.

Learning-based navigation, 2010s–present

The most recent turn is toward learning. Instead of hand-coding the rules that connect sensors to actions, researchers train policies—functions mapping sensor input to control output—from data5. Learning-based navigation draws on deep learning and reinforcement learning, the latter a framework in which a robot improves by trial and error against a reward signal6. Early successes included learned policies for drone flight and for legged locomotion over rough terrain, where hand-designed controllers had struggled to generalize. In navigation, learned methods have been used for terrain classification, for predicting the motion of people and vehicles, and increasingly for end-to-end control from camera images to motor commands.

Learning has not replaced the earlier approaches so much as added a layer. Probabilistic mapping and localization remain standard, because a learned policy still needs to know where it is and what surrounds it. The open questions are the ones the field has carried throughout: how to guarantee that a learned policy is safe, how to transfer what a robot learns in simulation to the real world, and how to combine the flexibility of learning with the reliability of explicit models. The history of mobile robotics has been a long negotiation between telling the robot about the world and letting it find out—and that negotiation is not finished.

References

  1. Mobile robot (introduction, History), Wikipedia
  2. DARPA Grand Challenge (History and background, introduction), Wikipedia
  3. Behavior-based robotics (History), Wikipedia
  4. Simultaneous localization and mapping, Wikipedia
  5. Robot learning, Wikipedia
  6. Reinforcement learning, Wikipedia

The frameworks of Mobile Robotics

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1940-1975

    Teleoperation

    How can a person safely control a machine at a distance?

    Fly to its tower in Nousfarer

  2. 1968-1995

    Deliberative robotics

    How should a robot plan its actions before moving?

    Fly to its tower in Nousfarer

  3. 1985-2005

    Reactive control

    How can a robot respond to its sensors in real time?

    Fly to its tower in Nousfarer

  4. 1986-Present

    Artificial Potential Fields

    How can a robot steer around obstacles without searching?

    Fly to its tower in Nousfarer

  5. 1995-Present

    Probabilistic robotics

    How can a robot act when it is uncertain where it is?

    Fly to its tower in Nousfarer

  6. 2004-2015

    Field and challenge-driven autonomy

    Can robots complete real courses and city streets on their own?

    Fly to its tower in Nousfarer

  7. 2013-Present

    Robot learning

    Can a robot learn to navigate from experience instead of rules?

    Fly to its tower in Nousfarer

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

TeleoperationDeliberative roboticsReactive controlArtificial Potential FieldsProbabilistic roboticsField and challenge-driven autonomyRobot learning

Beyond the field

Where Mobile Robotics's ideas came from in other fields, and what they led to there: the strongest 12 of 26 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Mobile Robotics is a planet in Nousfarer

Its city has a tower for each of these 7 frameworks. Land, walk in, and climb one concept at a time.

Land on Mobile Robotics

Related fields

Fields its ideas are linked to

More in Robotics

See where Mobile Robotics sits on the Map of Human Knowledge