Do all robots use AI?

No. Many rely on programmed sequences and conventional control, while others incorporate learned components.

Must a robot be autonomous?

Robotics includes different degrees of autonomy, including remote operation.

Is a chatbot a physical robot?

A software-only conversational system is not physical robotics, though it may provide an interface to a robot.

AI Glossary / Reinforcement learning and robotics / Intermediate

Robotics

Definition

Robotics is the field concerned with designing, building, programming, controlling, and using robots. It brings together mechanical systems, sensing, computation, and physical action. Artificial intelligence can support perception, planning, and learning within robotics, but a robot's usefulness also depends on its mechanics, control, integration, and ability to operate reliably in its intended environment.

What it means

Robotics connects software decisions with physical consequences. A robot may move through a building, position a tool, inspect equipment, or manipulate an object. Its actions are constrained by geometry, forces, power, sensing quality, and the surrounding world.
Robots vary widely in form and autonomy. An industrial arm, a mobile inspection platform, and a remotely operated underwater machine can all be robotics systems. There is no requirement that a robot resemble a person, understand conversation, or use a neural network.

How it works

A robotics system combines hardware and software around a task. Sensors measure relevant conditions. Computation interprets those measurements and selects actions. Actuators produce motion or apply forces, while feedback helps correct differences between desired and observed behavior.
Kinematics describes motion and geometric relationships, such as how joint positions determine a tool's location. Dynamics accounts for forces and motion. Planning selects feasible movements or task sequences, and control generates commands that make the physical system follow them.
Integration links these functions with calibration, timing, power, communication, and the operating workflow. A camera may identify an object correctly while a poorly calibrated coordinate transform sends the gripper to the wrong place. A mechanically suitable arm may still be ineffective if objects arrive in an uncontrolled orientation.
Testing should cover the complete task, including unsuccessful attempts and recovery. Simulation can support development, but contact, friction, wear, lighting, and unexpected objects can differ from the model. Reliable deployment requires understanding those differences rather than assuming a successful demonstration will repeat indefinitely.

Three practical examples

Manufacturing: use robotic manipulators to position, assemble, or inspect components within a defined process.
Logistics: combine mobile navigation and handling systems to move goods between work areas.
Field inspection: gather measurements in locations where remote operation or autonomous movement supports the inspection task.

Think of it this way

Think of robotics as organizing a body, senses, and a control process around a job. Better decision software can help, but it cannot make an unsuitable gripper securely hold every object or make a weak actuator lift an arbitrary load. The parts must work together for the task to succeed.

Worked example

Consider a fictional robot cell that transfers small boxes from a conveyor to trays. A nominal cycle uses 0.4 seconds to locate a box, 0.6 seconds to approach it, 0.3 seconds to grasp, 0.9 seconds to transfer, and 0.3 seconds to release. Assume these steps occur sequentially, with no overlap. Total nominal cycle time is 2.5 seconds, giving a theoretical maximum of 3600/2.5 = 1440 attempts per hour.
Now suppose the cell is available for 85% of the scheduled hour and 96% of attempts produce an acceptable transfer. Assume failed attempts consume the same cycle time and need no extra recovery time for this calculation. Expected acceptable transfers are 1440 times 0.85 times 0.96 = 1175.04 per hour. This is an expectation under simplified assumptions, not an exact production count.
An improved perception model reduces localization time from 0.4 to 0.2 seconds. The new nominal cycle is 2.3 seconds. Keeping the same availability and success assumptions gives (3600/2.3) times 0.85 times 0.96, approximately 1277.22 acceptable transfers per hour. The improvement is about 102.18 per hour, or 8.70%.
However, suppose the incoming conveyor supplies only 1100 boxes per hour. The cell cannot deliver 1277 boxes per hour when only 1100 arrive, even if every downstream component is capable. The upstream supply becomes the limiting factor in this simplified flow.
Finally, if failed grasps require additional recovery time, the earlier throughput estimates are optimistic because the equal-cycle assumption is false. Measuring the recovery distribution would be necessary to revise the calculation. This example shows why robotics performance must be assessed across sensing, motion, reliability, recovery, and the surrounding process rather than through one component's speed alone.

Illustrative case study

Fictional case: JuniperPackaging selects a robot arm using payload and reach specifications. During a pilot, flexible packages deform in the gripper and produce unreliable placement. The team changes the gripping approach, improves package presentation, and adds checks for successful pickup.
It then measures completed acceptable transfers and recovery time across representative package types. This invented case illustrates system integration; it does not claim a particular gripper design, robot brand, or AI model will solve the problem.

Common misconceptions and mistakes

Common mistakes include treating robotics as another name for AI, assuming more autonomy is always better, and evaluating only the robot's nominal movement speed. Payload, reach, repeatability, sensing, task variability, and recovery behavior all affect suitability. Software process automation is also distinct from physical robotics, even when its marketing uses the word robot.

Business use

Businesses should begin with a clearly defined physical workflow and measurable acceptance criteria. Assess the full operating cost, integration effort, maintenance, task variation, and interaction with existing equipment and people. A useful pilot demonstrates repeatable task completion under representative conditions and records the work needed when the robot cannot complete an attempt.

Check your understanding

What is the nominal cycle time in the example?

Answer: 2.5 seconds.

Why is expected acceptable output below 1440 per hour?

Answer: Availability and unsuccessful attempts reduce it.

Why does faster perception not guarantee higher final throughput?

Answer: Another part of the workflow, such as incoming supply, may become the bottleneck.

Related terms

Kinematics
Motion planning
Autonomy
Robot
actuator
sensor
dynamics
control
manipulation
navigation
human-robot interaction
system integration

Learn next

Next, map a simple robot task from sensing through action and verification. Identify the main physical constraints and measure complete cycle outcomes before deciding which component most needs improvement.

Sources

IEEE Spectrum ROBOTS, What Is a Robot?: https://robotsguide.com/learn/what-is-a-robot Kevin M. Lynch and Frank C. Park, Modern Robotics: Mechanics, Planning, and Control, author resource: https://hades.mech.northwestern.edu/index.php/Modern_Robotics

Previous: Robotic process automation Next: Robust AI