Autonomous Systems

Autonomous Systems
Systems that perform specified functions with varying degrees of human supervision
Category	Autonomous control and decision-making systems
Area	Robotics, control engineering, computer science and systems engineering
Core components	Task logic, sensing or information inputs, state estimation, planning, control and monitoring
Applications	Mobile robots, industrial vehicles, automated driving and remote inspection
Related concepts	Robotics, Physical AI, Embodied AI, Autonomous Navigation, Industrial Automation
Security considerations	Unauthorized control, compromised communications, software integrity and access management
Relevant standards	ISO 8373, SAE J3016, ISO 3691-4:2023

Autonomous systems are engineered systems capable of performing specified tasks or functions with some degree of independence from continuous human intervention. Depending on their design, such systems may monitor conditions, determine actions, adjust their behavior and respond to changes in their operating environment. The extent of their autonomy depends on the tasks assigned to them, the circumstances under which they operate and the responsibilities retained by human operators.

The term is used across several areas of engineering and computing, including robotics, artificial intelligence, control engineering, automated transport and distributed computing. In physical systems, autonomous operation commonly involves sensors, computing equipment, control software and mechanisms capable of acting upon the environment. In software systems, autonomy may instead concern the execution of computational tasks or decisions without direct human intervention at every step.

Autonomy should not be confused with unrestricted independent decision-making. A system may operate without continuous supervision while remaining subject to programmed objectives, operating limits, external instructions and safety constraints. Neither autonomy nor the use of artificial intelligence establishes that a system is reliable or safe in all circumstances.

Definitions and scope

Autonomy and automation

Automation is the use of technology to perform tasks or processes with reduced direct human involvement. Automated equipment may follow a predetermined sequence of instructions and respond to inputs according to established control rules.

Autonomy concerns the ability of a system to carry out assigned functions while exercising some independence in its operation. This can involve selecting among alternative actions, responding to conditions not specified as individual commands or adjusting a plan as circumstances change.

The distinction is not absolute. Automated systems can incorporate autonomous functions, and autonomous systems commonly depend on automated control processes. The relevant question is which decisions and actions the system can perform independently within its specified operating conditions.

Autonomy and artificial intelligence

Autonomous systems are not necessarily artificial intelligence systems.

Conventional feedback controllers, programmed decision rules and established planning algorithms can support autonomous operation. Machine learning provides additional methods for selected perception, prediction and decision-making tasks, but it is not a defining requirement of autonomy.

Similarly, artificial intelligence systems do not necessarily control physical equipment or operate autonomously. An AI model may provide information to a human operator without independently executing any consequential action.

Relationship to robotics

Robotics concerns the engineering and operation of robots, including their mechanical design, sensing, motion and control.

An autonomous robot is a robot capable of performing specified functions with some independence from direct human control. Other robots may operate through teleoperation or closely supervised control.

The robotics vocabulary standard ISO 8373:2021 provides terminology for robots and robotic systems. Its scope is robotics rather than a universal definition of autonomy across all engineering domains.

Degrees and dimensions of autonomy

Autonomy is often described in terms of the functions a system performs without human intervention. However, systems operating in different environments cannot always be compared meaningfully using a single numerical scale.

Researchers at the National Institute of Standards and Technology (NIST) investigated this problem through the Autonomy Levels for Unmanned Systems (ALFUS) framework.

The framework considers three important dimensions:

• Task complexity: The difficulty and range of tasks a system is expected to perform.
• Environmental difficulty: The complexity and variability of the environment in which the system operates.
• Human interaction: The nature and extent of human involvement required during operation.

These dimensions provide a way to characterize autonomous capability in relation to a particular mission or application. They do not establish a universally applicable autonomy score for all technologies.

Human involvement

Human participation can vary during the operation of an autonomous system.

In manual or direct control, an operator determines the principal actions carried out by a machine.

In supervisory control, a system performs delegated functions while a human monitors its operation and may intervene when necessary.

In autonomous execution, the system carries out an assigned task without requiring continuous individual commands from a human operator.

These descriptions identify broad forms of operation rather than universally standardized autonomy levels. A single system may support more than one operating mode.

Driving automation levels

Road-vehicle automation has a separate, domain-specific classification.

SAE International's SAE J3016 provides a taxonomy for driving automation systems used in on-road motor vehicles. The September 2026 revision continues to distinguish levels from 0 through 5:

Level	Classification	General distinction
0	No driving automation	Human driving remains necessary
1	Driver assistance	The system supports either steering or speed control while the driver supervises
2	Driver support	The system supports both steering and speed control while the driver supervises
3	Conditional driving automation	Automated driving under defined conditions, with a human required to resume driving when necessary
4	High driving automation	Automated driving within defined conditions without requiring a human to mitigate driving risks
5	Full driving automation	Automated driving under all conditions in which a human could drive

This classification addresses the performance of the dynamic driving task and the respective responsibilities of the human driver and the automation system. It should not be applied directly to warehouse robots, industrial controllers, spacecraft or software agents.

System architecture

The architecture of an autonomous system depends on its purpose, environment and performance requirements.

Physical autonomous systems commonly combine information acquisition, state estimation, planning, control and actuation. NIST's measurement-science research describes robotic systems in terms of perceiving their surroundings and current state, planning actions and executing those actions through physical mechanisms.

Sensing and information acquisition

Sensors provide measurements about a system and its environment.

Depending on the application, these can include cameras, distance sensors, inertial measurement units, positioning systems, encoders, force sensors and other measurement devices.

A system may also receive information through network communications or external data sources.

Sensor measurements are subject to uncertainty, limited range, interference and other constraints. The reliability of autonomous behavior therefore depends partly on the quality and interpretation of available information.

State estimation

State estimation concerns determining the current condition of a system from available measurements and models.

For a mobile robot, the relevant state may include its position, orientation and velocity. Other applications may require estimates of equipment condition, environmental features or the location of objects.

When measurements are incomplete or uncertain, estimation methods can combine multiple observations and model assumptions to produce an estimate of the system's state.

Probabilistic approaches, including Bayesian filtering methods, are important in robotic localization and related estimation problems.

Perception and environmental understanding

Computer vision and other perception methods can be used to interpret information acquired from sensors.

Tasks may include identifying obstacles, locating objects, estimating movement or producing a representation of the operating environment.

Sensor fusion can combine different measurement sources. Its performance depends on sensor characteristics, estimation methods and the assumptions used to combine information.

Planning and decision-making

Planning determines a sequence of actions or a policy for achieving a specified objective while accounting for relevant constraints.

A planning problem can involve finding a route, scheduling operations, selecting a manipulation sequence or determining an appropriate response to an environmental condition.

Algorithms may operate with complete information, incomplete information or uncertainty about future events.

Decision-theoretic planning provides mathematical approaches for selecting actions when outcomes are uncertain.

Control and actuation

A controller generates commands intended to produce a desired system response.

Physical autonomous systems may use feedback control to regulate motion, position, velocity or force. Actuators then translate control signals into physical effects.

Control performance depends on properties of the physical system, including dynamics, actuator limitations, disturbances and feedback quality.

Autonomous planning and low-level control are related but distinct functions. A planner may select a desired trajectory while a controller attempts to follow it using available measurements.

Monitoring and supervision

Monitoring functions evaluate system operation and may detect events requiring intervention.

Depending on the design, the system may report its condition, request operator assistance or transition to a predefined response when a fault or operating limit is detected.

The appropriate monitoring and intervention mechanisms depend on the hazards, operational requirements and available communication or control interfaces.

Decision and control processes

Autonomous systems may combine several forms of decision-making.

Rule-based behavior

Rule-based systems select actions according to explicitly specified conditions.

For example, a controller may change its operating state when a measured quantity crosses a configured threshold.

Such behavior can be part of an autonomous system, but a fixed rule does not necessarily provide the flexibility needed for complex or uncertain environments.

Model-based planning

Model-based methods use representations of the system, environment or possible outcomes to determine actions.

These methods may include graph search, optimization, trajectory planning and other established algorithms.

Their effectiveness depends on the accuracy of the model, the assumptions of the planning method and the system's ability to respond to discrepancies between predictions and observations.

Decision-making under uncertainty

A system may need to act when it cannot directly observe all relevant environmental conditions.

Markov decision processes and partially observable Markov decision processes provide mathematical models for sequential decision-making, including cases involving uncertain transitions or incomplete observations.

These models are used in robotics and artificial intelligence research. Their practical application depends on computational feasibility and the availability of suitable models.

Machine learning

Machine learning can be used in autonomous systems for tasks such as perception, prediction or policy development.

Reinforcement learning addresses problems in which a learning agent interacts with an environment and develops behavior according to a specified reward formulation.

Learned components do not eliminate the need to evaluate system performance under relevant conditions.

A learned model may perform successfully during training or testing while exhibiting different behavior when deployed under changed conditions. Its integration into a physical system therefore requires appropriate evaluation and monitoring.

Operating conditions and limitations

Autonomous capability must be understood in relation to the conditions for which a system is designed.

An autonomous mobile robot may be able to navigate a mapped facility but may not be capable of operating reliably in an unfamiliar outdoor environment.

Similarly, an automated driving function may be designed for particular road, traffic, speed or environmental conditions.

In the automated-driving domain, the term operational design domain (ODD) describes the operating conditions under which a driving automation system or feature is designed to function. The concept is central to the SAE J3016 taxonomy.

For other autonomous applications, operational boundaries may be expressed through system requirements, environmental specifications, operating procedures or application-specific terminology.

Important limitations can include:

• Incomplete or inaccurate sensor information.
• Environmental conditions outside the intended operating range.
• Limited computational resources or communication availability.
• Mechanical and actuator constraints.
• Incorrect assumptions in models or planning methods.
• Software faults and unexpected interactions between system components.

These limitations are relevant when interpreting manufacturer claims, published performance results and statements about a system's degree of autonomy.

Types and applications

Autonomous mobile robots

Autonomous mobile robots are designed to perform movement-related tasks with varying degrees of independence from direct control.

Their functions can involve localization, navigation, obstacle detection and path planning.

Industrial applications include selected material-handling and transport operations. Safety requirements for driverless industrial trucks and their systems are specified in ISO 3691-4:2023, which covers relevant vehicle categories including automated guided vehicles and certain autonomous mobile robots.

Automated driving systems

Automated driving systems perform portions or all of the dynamic driving task within the conditions associated with their design and automation level.

Their operation can involve environmental perception, localization, trajectory planning, motion control and interactions with other road users.

SAE J3016 distinguishes driver-support features from automated driving systems capable of performing the complete dynamic driving task under specified conditions.

Industrial automation

Autonomous capabilities can be incorporated into manufacturing and logistics systems.

Examples include equipment capable of planning movements, adapting to measured operating conditions or performing selected tasks with supervisory rather than continuous manual control.

The presence of automatic machinery does not establish that an entire production system is autonomous. The relevant capabilities must be assessed at the level of the equipment, task or integrated system.

Inspection and remote operations

Robotic inspection platforms may incorporate autonomous movement, data acquisition and monitoring functions.

These systems may also operate through teleoperation or combinations of human supervision and autonomous functions.

The level of autonomy depends on the particular system and the conditions under which it is deployed.

Software-based autonomous systems

Autonomy is not restricted to physical machines.

Software systems may execute delegated tasks, select actions based on information and respond to changing computational conditions.

However, autonomous execution does not necessarily imply independent judgment, general intelligence or the ability to operate outside predefined permissions and system constraints.

Human oversight and interaction

Human oversight is an important design and operational consideration in autonomous systems.

Depending on the application, operators may assign tasks, monitor performance, approve certain actions, manage faults or assume direct control.

The human-machine interface must be considered together with the capabilities and limitations of the autonomous functions.

NIST's ALFUS research treats human interaction as one dimension in assessing autonomy. This recognizes that the amount and type of human involvement are relevant to a system's actual operating capability.

Intervention arrangements must also account for the particular environment. A remote command, for example, cannot be assumed to provide immediate physical intervention under all network conditions.

Safety engineering

Autonomous systems capable of influencing physical processes may create hazards through unexpected actions, perception errors, control failures or inappropriate responses to environmental changes.

Safety engineering involves identifying hazards, evaluating risks and implementing measures appropriate to the intended application.

Depending on the system, measures may include:

• Defining permitted operating conditions.
• Establishing relevant safety requirements.
• Testing responses to faults and abnormal conditions.
• Restricting or stopping hazardous motion.
• Separating safety-related control functions where required.
• Providing suitable operator information and intervention procedures.
• Evaluating the complete integrated system.

These measures are not interchangeable, and the appropriate combination depends on the hazards and applicable standards.

Application-specific standards

There is no single safety standard that covers every autonomous system.

For example, ISO 3691-4:2023 specifies safety requirements and means of verification for driverless industrial trucks and their systems.

For road vehicles, ISO 21448:2022 addresses safety of the intended functionality. It considers hazards arising from insufficient specification or performance of intended functions, especially those involving environmental sensing and situational awareness.

ISO 21448 does not replace requirements concerning hardware and software faults addressed by other automotive standards, nor does it directly cover cybersecurity threats.

Safety requirements must therefore be selected according to the system's function, environment and applicable regulatory context.

Verification, validation and performance measurement

Verification concerns whether a system or component satisfies specified requirements.

Validation addresses whether the system is suitable for its intended use and operating conditions.

Evaluating autonomy requires more than demonstrating a successful task on one occasion.

Relevant assessment methods may include controlled experiments, performance measurements, simulation, scenario testing, physical trials and examination of responses to failures.

The appropriate evaluation depends on the system's tasks and the environment in which it is expected to operate.

NIST's Measurement Science for Robotics and Autonomous Systems Program develops performance metrics, test methods, datasets and related measurement approaches for evaluating robotic systems.

The program emphasizes assessing both individual functions and the performance of the integrated system.

Reliability and uncertainty

Real-world performance can depend on changing environmental conditions, imperfect measurements and disturbances.

Testing must therefore consider whether observed results are representative of the intended operating conditions.

A system's performance in one environment should not automatically be generalized to every environment or application.

Cybersecurity and system identity

Connected autonomous systems may exchange commands, software updates, operational data and status information with external services or operators.

These connections introduce risks such as unauthorized access, compromised communications and modification of software or operational instructions.

Cybersecurity controls can include:

• Authentication of devices, services and users.
• Authorization of commands and access to resources.
• Protection of communications.
• Integrity controls for software and updates.
• Monitoring and response to security incidents.
• Appropriate separation of network and control functions.

NIST SP 800-82 Revision 3 provides guidance on securing operational technology while accounting for its performance, reliability and safety requirements.

Machine identity and authorization

Machine identity mechanisms can help identify and authenticate controllers, machines and services participating in a connected autonomous system.

Authentication establishes a basis for recognizing the communicating entity. Authorization separately determines which operations that entity may perform.

A valid credential does not guarantee that software is uncompromised, that a command is safe or that the equipment is operating correctly.

Identity management is therefore one component of a broader cybersecurity and system-safety approach.

Artificial intelligence risk management

Some autonomous systems use artificial intelligence for perception, decision-making or other functions.

For these systems, AI-specific risks may arise from limitations in model performance, data quality, robustness or the circumstances in which an AI component is deployed.

The NIST Artificial Intelligence Risk Management Framework (AI RMF 1.0) was published in January 2023. It provides voluntary guidance for organizations involved in the design, development, deployment and use of AI systems.

The framework organizes AI risk-management activities around four functions:

• Govern: Establish and maintain organizational structures and processes for AI risk management.
• Map: Understand the context, intended purpose and risks of an AI system.
• Measure: Assess relevant system characteristics and risks using appropriate methods.
• Manage: Prioritize and respond to identified risks throughout the system lifecycle.

These functions are intended to support ongoing risk management rather than prescribe a single engineering development sequence.

The AI RMF is not an autonomous-system safety certification and does not replace the engineering standards applicable to physical equipment.

Research challenges

Research on autonomous systems addresses problems involving uncertain information, changing environments, interaction with people and the reliable execution of complex tasks.

Relevant areas include:

• Perception and estimation: Obtaining reliable information under uncertain sensing conditions.
• Planning: Selecting feasible actions while respecting system and environmental constraints.
• Robust control: Maintaining satisfactory behavior despite disturbances and modeling errors.
• Human interaction: Defining suitable arrangements for supervision, communication and intervention.
• Verification and testing: Establishing evidence that systems satisfy requirements under specified conditions.
• Learning-enabled autonomy: Evaluating the behavior of systems incorporating learned models.
• System integration: Understanding how the performance of individual components affects the complete system.

These subjects are treated in the robotics, planning, control and artificial intelligence literature.

Progress in one technical component should not automatically be interpreted as an equivalent improvement in the autonomy or safety of an entire deployed system.

Selected academic books

• LaValle, Steven M. (2006). Planning Algorithms. Cambridge University Press. ISBN 978-0-521-86205-9. Covers motion planning, planning under uncertainty, decision theory and control-related planning.
• Thrun, Sebastian, Wolfram Burgard, and Dieter Fox (2005). Probabilistic Robotics. MIT Press. ISBN 978-0-262-20162-9. Covers probabilistic state estimation, localization, mapping and decision-making under uncertainty.
• Lynch, Kevin M., and Frank C. Park (2017). Modern Robotics: Mechanics, Planning, and Control. Cambridge University Press. ISBN 978-1-107-15630-2. Covers robot mechanics, trajectory generation, planning and control.
• Russell, Stuart J., and Peter Norvig (2020). Artificial Intelligence: A Modern Approach. 4th edition. Pearson. ISBN 978-0-13-461099-3. Covers intelligent agents, problem solving, planning, uncertainty and artificial intelligence.
• Sutton, Richard S., and Andrew G. Barto (2018). Reinforcement Learning: An Introduction. 2nd edition. MIT Press. ISBN 978-0-262-03924-6. Covers foundational theory and algorithms for reinforcement learning.

Standards and technical guidance

• NIST — Autonomy Measures for Robots
• NIST — A Framework for Autonomy Levels for Unmanned Systems
• NIST — Measurement Science for Robotics and Autonomous Systems
• ISO 8373:2021 — Robotics Vocabulary
• ISO 3691-4:2023 — Driverless Industrial Trucks and Their Systems
• ISO 21448:2022 — Safety of the Intended Functionality
• SAE J3016 — Driving Automation Taxonomy
• NIST AI RMF 1.0
• NIST SP 800-82 Revision 3 — Guide to Operational Technology Security

See also

• Physical AI
• Robotics
• Embodied AI
• Artificial Intelligence
• Industrial Automation
• Autonomous Navigation
• Autonomous Mobile Robots
• Computer Vision
• Sensor Fusion
• Machine Identity
• Robot Authentication
• Robot Cybersecurity
• AI Governance
• Robot Safety Risk Assessment