Physical AI Fundamentals

Physical AI Fundamentals
Foundational concepts for understanding AI-enabled systems in the physical world
Page type	Educational overview and reading guide
Principal subjects	Robotics, Artificial Intelligence, control engineering and physical systems
Core concepts	Sensing, perception, state estimation, planning, control and feedback
Advanced topics	Embodied AI, robot learning, Autonomous Systems and simulation
Trust and safety	Machine Identity, Robot Cybersecurity, risk assessment and evaluation
Reference material	Robotics textbooks, published research, NIST guidance and ISO standards

Physical AI fundamentals is an introductory guide to the scientific and engineering concepts used to understand physical artificial intelligence (physical AI). It brings together principles from robotics, artificial intelligence, mechanical engineering, sensing, control theory, computer science and system safety. The guide is intended to help readers understand how AI-enabled systems obtain information about physical environments, make decisions and, when equipped with suitable mechanisms, perform physical actions.

The expression physical AI is used for applications of artificial intelligence involving the physical world, including intelligent robots and systems that interpret information from physical environments. It does not denote a single standardized technical architecture. The underlying systems may use different combinations of conventional control algorithms, statistical models and machine learning. NIST's robotics research describes perception, planning, action and system-level evaluation as important capabilities of advanced robotic systems.

This page is a learning guide rather than a separate scientific discipline. Its purpose is to introduce established concepts, explain how they relate to one another and direct readers to more specialized articles and academic sources.

Scope and essential concepts

Physical AI, robotics and automation

Robotics is an engineering and scientific field concerned with robots, including their mechanisms, sensing, motion, control and applications. It covers robotic systems whether or not they use machine learning.

Artificial intelligence includes computational methods for tasks such as perception, reasoning, learning and decision-making. An AI system does not necessarily control a physical machine.

Automation involves using technology to execute processes or functions with reduced direct human intervention. Many automated systems operate through programmed logic and feedback control without requiring machine learning.

Autonomous systems can perform specified functions with varying degrees of independence from continuous human control. Their capabilities depend on the task, operating environment and system design.

The concepts overlap, but they are not interchangeable. A programmed industrial manipulator can be a robot without using a learned policy. An image-recognition model may be an AI component without being a robot. A mobile robot using machine learning for perception and conventional control for movement can incorporate both approaches.

The international standard ISO 8373:2021 provides terminology relating to robotics. It should not be interpreted as establishing a universal definition or certification of physical AI.

The interaction between information and physical action

An AI-enabled physical system may need to acquire measurements, estimate relevant conditions, select an action, execute it and evaluate the results.

For introductory purposes, these activities can be represented as a general sequence:

1. Observe: Acquire information from sensors or other available inputs.
2. Estimate: Infer the current condition of the robot, task or environment.
3. Plan or decide: Select an action consistent with the assigned objective and constraints.
4. Act: Produce the intended response through a controller or physical mechanism.
5. Monitor: Use feedback to assess the response and determine whether adjustment is needed.

This sequence is an educational explanation, not a claim that every physical AI system uses the same software architecture. Practical systems may perform these operations simultaneously, repeatedly or through several interacting controllers.

1. Physical systems and robot mechanics

Understanding the physical structure of a machine is a prerequisite for understanding how it can interact with its surroundings.

Mechanical structure

Robotic mechanisms consist of physical components such as links, joints, frames, wheels, tracks or other movable structures.

A manipulator may contain several connected links and joints, while a mobile robot may combine a locomotion platform with additional sensing or manipulation equipment.

The mechanical arrangement determines which movements are possible and affects the system's reach, payload, speed, stiffness and other performance characteristics.

Degrees of freedom

A degree of freedom describes an independent coordinate needed to specify the configuration of a mechanical system.

For robotic mechanisms, the number of degrees of freedom depends on the arrangement of joints and the constraints acting on the system.

This concept is important when describing movement and determining whether a robot can reach a particular configuration or orientation.

Coordinate frames and rigid-body motion

Robotics uses coordinate systems to describe positions, orientations and movements.

A coordinate frame provides a reference for expressing the location or orientation of objects. Different frames may describe the robot's base, a sensor, an end effector or a feature of the environment.

Rigid-body transformations relate measurements and movements expressed in different coordinate frames.

These mathematical tools are fundamental to robot kinematics, perception and motion planning.

Kinematics and dynamics

Kinematics describes motion without directly analyzing the forces causing it.

Forward kinematics determines a robot component's pose from its joint configuration. Inverse kinematics addresses the problem of determining configurations that achieve a required pose.

Dynamics concerns the relationship between motion and quantities such as forces, torques, mass and inertia.

Both subjects are central to mechanical modeling and control.

Actuators and mechanical limits

Actuators convert supplied energy into mechanical force or movement.

Robotic applications may use electric, hydraulic or pneumatic actuation, depending on their performance requirements.

Physical systems are constrained by actuator capabilities, joint limits, mechanical loads and environmental interactions. A software command does not guarantee that a requested movement is physically achievable.

Continue reading: Robotics, Robot Kinematics, Robot Dynamics, Actuators and Robot Manipulation.

2. Sensors and perception

Physical AI systems obtain information about the world through measurements. The available information depends on the sensors, the environment and the algorithms used to interpret observations.

Sensor types

Robotic sensing can include:

• Cameras for acquiring visual information.
• Distance and ranging sensors for measuring aspects of surrounding geometry.
• Position encoders for observing joint or motor movement.
• Inertial sensors for measuring acceleration or angular motion.
• Force and torque sensors for measuring mechanical interactions.
• Tactile sensors for selected contact and manipulation tasks.

Different sensors provide different measurements and have different limitations. Appropriate sensor selection depends on the information required by the application.

Computer vision

Computer vision concerns computational methods for interpreting visual information.

In robotic systems, such methods may support object recognition, feature detection, pose estimation and other perception tasks.

Some approaches rely on geometric or statistical models, while others use machine learning. The use of a camera does not, by itself, establish that a system is using artificial intelligence.

NIST investigates methods for assessing the performance of robotic perception systems, including sensing technologies and both conventional and AI-based algorithms.

Sensor fusion

Sensor fusion combines measurements from more than one source to estimate information relevant to a task.

For example, a localization system may use different sources of motion and environmental information.

The usefulness of combined measurements depends on calibration, measurement uncertainty, the estimation method and the assumptions about the sensors involved.

State estimation

A robot cannot necessarily measure every variable needed for control directly.

State estimation uses available observations and models to estimate quantities such as position, velocity, orientation or other relevant conditions.

Probabilistic robotics studies methods that explicitly account for uncertainty in sensing, estimation and decision-making.

Localization and mapping

Localization concerns estimating a robot's pose with respect to a reference frame or map.

Mapping concerns constructing a representation of an environment from available observations.

Simultaneous localization and mapping (SLAM) addresses the coupled problem of estimating the robot's pose while constructing or updating a map.

Continue reading: Computer Vision, Sensors, Sensor Fusion, Robot Localization and Simultaneous Localization and Mapping.

3. Planning, decision-making and control

Sensing and perception provide information that can support the selection and execution of actions.

Planning

Planning concerns determining actions or sequences of actions that satisfy an objective under relevant constraints.

A robotic planning problem may require finding a route around obstacles, selecting an object-manipulation sequence or generating a feasible trajectory.

Planning methods differ according to the available information, the representation of the environment and the constraints imposed by the task.

Motion planning

Motion planning studies how to move a robot between configurations or states while satisfying the requirements of a task.

Depending on the system, constraints may involve collision avoidance, joint limits, vehicle steering, motion continuity or other physical limitations.

A planned path describes a geometric route through a configuration or physical space. A trajectory additionally specifies motion over time.

Feedback control

Feedback control uses measurements of actual system behavior to adjust control commands.

A controller may compare a desired value with a measured value and modify the actuator input accordingly.

Examples of controlled quantities include position, velocity, force and torque.

Robot control methods range from established feedback algorithms to model-based approaches. The selection of an appropriate controller depends on the system's dynamics and operating requirements.

Planning versus control

Planning and control are related but distinct activities.

A planning component may determine where a robot should move, while a controller produces the commands needed to follow the selected motion.

An integrated system must consider the consistency of its planned actions with the physical capabilities of the robot.

Decision-making under uncertainty

An autonomous system may need to choose an action without having complete information about its environment.

Mathematical frameworks such as Markov decision processes and partially observable Markov decision processes are used to study sequential decision-making, including cases involving uncertain outcomes or incomplete observations.

Continue reading: Motion Planning, Feedback Control, Autonomous Navigation, Autonomous Systems and Robot Control.

4. Learning and artificial intelligence

Learning-based methods provide ways to develop models or policies from data. They may be used in physical systems, but they are not required for every robotic task.

Supervised learning

In supervised learning, a model is trained using data that includes target outputs or labels.

In robotics, supervised methods may support selected perception and prediction tasks.

The performance of such models depends on their training data, evaluation methods and conditions encountered during deployment.

Learning from demonstrations

Robot learning can use recorded demonstrations to develop behavior related to a task.

Demonstrations may include observations and actions obtained from operators or other systems.

Learning from demonstrations does not remove the need to establish whether the resulting behavior is suitable for the intended physical environment.

Reinforcement learning

Reinforcement learning studies methods in which an agent learns through interactions with an environment and reward-based feedback.

The learning problem concerns selecting behavior that improves expected cumulative reward under the assumed formulation.

Reinforcement learning provides a family of algorithms and mathematical models; it is not a guarantee that a trained policy will behave correctly outside the conditions considered during development.

Simulation and transfer to real systems

Simulation provides a controlled environment for studying robot models, developing control methods and evaluating selected behaviors.

However, simulated conditions may differ from physical operation because of modeling errors, sensor characteristics, contact behavior and environmental variation.

A successful simulation result is therefore not sufficient evidence of reliable performance in a real deployment.

Robotic evaluation must account for the intended application and the operating conditions in which the system will be used.

Vision-language-action research

Some contemporary robotics research investigates models that combine visual observations, natural-language information and robot actions.

The RT-2 research, presented at the Conference on Robot Learning in 2023, studied vision-language-action models that represent robot actions using token sequences and combine robotic training data with vision-language pretraining.

This is an example of a published research approach, not evidence that all physical AI systems require language models or use the same architecture.

Continue reading: Machine Learning, Embodied AI, Reinforcement Learning, Robot Learning and Robot Foundation Models.

5. System integration and reliable operation

Robotic systems combine several interacting technical components.

A system may contain mechanical actuators, embedded computers, perception modules, planning software, controllers, communication interfaces and monitoring equipment.

The performance of the complete system depends on the behavior of these components and their interactions.

NIST emphasizes that robotic performance must be assessed both at the component level and at the level of the integrated system.

Timing and computation

Physical control systems must account for the time required to acquire measurements, process information and issue commands.

Communication delays, computing limitations and control update rates can influence the behavior of an integrated system.

The relevant timing requirements depend on the system's dynamics and intended tasks.

Physical and environmental constraints

Physical behavior is influenced by factors such as gravity, friction, contact, inertia and mechanical limits.

Environmental characteristics can also influence sensor measurements, navigation and manipulation.

Models and experiments must therefore be interpreted in relation to the assumptions and conditions under which they were developed.

Human interaction

Some robotic systems operate near people or directly interact with human users.

Human-robot interaction may involve shared workspaces, operator instructions, supervisory control or physical assistance.

The appropriate human involvement, interface design and safety measures depend on the application and its hazards.

6. Identity, authentication and cybersecurity

AI-enabled physical systems may interact with remote services, management systems, operators or other machines.

These interactions create security requirements distinct from perception, planning and physical control.

Identification and authentication

Machine identity concerns identifiers and credentials associated with non-human computing entities, including devices, services and workloads.

Identification distinguishes an entity within a particular system or namespace.

Authentication verifies a claim about an entity's identity using appropriate evidence.

Authorization determines which actions or resources are permitted to that entity.

These concepts are related but not interchangeable. A device identifier does not, by itself, establish that a device is genuine. Similarly, successful authentication does not establish that every command from the authenticated entity is authorized or safe.

NIST's zero trust guidance separates authentication from authorization and emphasizes access decisions appropriate to the resource and context.

Cryptographic credentials

Cryptographic keys and digital certificates can be used to support authentication and the protection of communications.

Secure key management includes considerations such as generation, storage, use, replacement and compromise response.

NIST SP 800-57 Part 1 Revision 5 provides guidance for managing cryptographic keying material.

Cybersecurity and physical consequences

A connected robot or control system may be exposed to risks involving unauthorized access, compromised communications or modification of software and commands.

NIST SP 800-82 Revision 3 addresses security for operational technology, including systems that monitor or interact with physical processes. It emphasizes security approaches that account for operational performance, reliability and safety.

Cybersecurity mechanisms do not replace mechanical safety or functional safety controls.

Continue reading: Machine Identity, Physical AI Identity, Robot Authentication, Public Key Infrastructure and Robot Cybersecurity.

7. Safety, evaluation and governance

Physical AI systems require evaluation in relation to their intended uses and the possible consequences of incorrect or unexpected behavior.

Safety engineering

Robotic systems may create hazards through unexpected motion, collision, inadequate control, mechanical faults or unsafe interaction with people and equipment.

Risk assessment identifies relevant hazards and supports the selection of appropriate protective measures.

For industrial robotics, the ISO 10218 series provides important safety requirements:

• ISO 10218-1:2025 addresses safety requirements for industrial robots.
• ISO 10218-2:2025 addresses safety requirements for industrial robot applications and robot cells.

Their requirements have defined scopes and must not be assumed to apply to every category of robot, including all medical, service or consumer systems.

Verification and validation

Verification examines whether a system or component meets specified requirements.

Validation examines whether the system is suitable for its intended purpose and operating conditions.

Evaluation may include measurements of accuracy, task performance, robustness, failure behavior and other application-specific characteristics.

NIST's robotics research emphasizes performance metrics, repeatable test methods and evaluation of integrated systems rather than relying only on broad statements about capability.

AI risk management

The NIST Artificial Intelligence Risk Management Framework (AI RMF 1.0), published in January 2023, provides voluntary guidance for organizations developing, deploying or using AI systems.

Its core functions are:

• Govern — establish organizational arrangements for AI risk management.
• Map — identify the context and relevant risks associated with AI systems.
• Measure — assess relevant characteristics and risks.
• Manage — prioritize and respond to identified risks.

These functions support continuing risk management rather than a mandatory development sequence.

The AI RMF is not, by itself, a robotics safety certification or proof that a system can operate safely in all physical environments.

Continue reading: AI Safety, Robot Safety, AI Governance, Robot Cybersecurity and Robot Safety Risk Assessment.

Example: understanding an autonomous mobile robot

The following is an illustrative learning example, not a description or performance assessment of a particular commercial product.

Consider a mobile robotic platform assigned to travel between two locations in a controlled indoor environment.

System function	Fundamental concept	Question for the learner
Measure its motion	Sensors and state estimation	How does the machine estimate its current position?
Observe its surroundings	Robot perception	What environmental information is available and how reliable is it?
Determine a route	Motion planning	Which routes satisfy the robot's constraints?
Execute movement	Actuation and feedback control	How are motor commands adjusted during movement?
Respond to detected hazards	Safety-related system behavior	What protective measures are required for the application?
Communicate with management systems	Authentication and authorization	Which entities may issue commands or access data?
Assess task performance	Verification and validation	What measurements demonstrate operation under the intended conditions?

This example shows how robotics, estimation, planning, control, cybersecurity and evaluation can be relevant to a single integrated system. The precise design and applicable requirements depend on the equipment and operational context.

Suggested learning pathway

The following sequence is an editorial study recommendation, not a formally standardized curriculum.

Stage	Topics to learn	Recommended reference
1. Foundations	Mathematics, physics, coordinate frames and robot mechanisms	Modern Robotics
2. Sensing	Sensors, measurement uncertainty and state estimation	Probabilistic Robotics
3. Motion and control	Kinematics, dynamics, trajectories and feedback	Modern Robotics
4. Planning	Search methods, motion planning and uncertainty	Planning Algorithms
5. Learning	Supervised learning, demonstrations and reinforcement learning	Reinforcement Learning: An Introduction
6. Integration	Architecture, physical constraints, testing and performance measurement	Springer Handbook of Robotics and NIST robotics research
7. Trust and safety	Identity, cybersecurity, risk management and applicable standards	NIST SP 800-207, NIST SP 800-82 and ISO 10218

Questions for self-study

After reading the introductory sections, a learner should be able to investigate the following questions:

1. What distinguishes a robotic system from a software-only AI model?
2. How do sensors, perception algorithms and state estimation differ?
3. What is the difference between kinematics and dynamics?
4. Why does motion planning not guarantee accurate physical execution?
5. How do conventional controllers differ from learned policies?
6. Why must results obtained in simulation be evaluated before physical deployment?
7. What is the distinction between identification, authentication and authorization?
8. Why do reliable operation, cybersecurity and functional safety require separate forms of evaluation?

These questions are intended as a reading aid. They are not a substitute for practical training or application-specific safety procedures.

Common misunderstandings

All robots require artificial intelligence

Robots may operate through programmed rules, conventional planning and feedback control. Machine learning is one possible component, not an essential feature of every robotic system.

Autonomous operation means unlimited independence

Autonomous capabilities are defined relative to tasks, operating conditions and human involvement. A system may perform one task autonomously while requiring supervision for other activities.

Successful simulation proves real-world performance

Simulation results depend on the models and assumptions used. Evaluation under the relevant physical operating conditions remains important.

Authentication makes a robot safe

Authentication helps verify the identity associated with a credential. It does not establish correct physical behavior, mechanical integrity or compliance with safety requirements.

A capable AI model is equivalent to a complete robotic system

A physical robotic system must integrate computing, sensing, mechanisms, control and other application-specific functions. The performance of one model cannot automatically be treated as the performance of the complete machine.

Glossary

Term	Meaning in this guide
Actuator	A component that converts supplied energy into mechanical movement or force
Autonomy	A system's capacity to perform specified functions with some degree of independence from direct human control
Control	Methods used to influence the behavior of a dynamic system
Feedback	Information about observed system behavior used to adjust subsequent operation
Kinematics	The description of motion without directly considering the forces producing it
Dynamics	The study of motion in relation to forces and torques
Localization	Estimating an entity's position and orientation relative to a reference
Perception	Interpreting sensor information to estimate relevant properties of the environment or system
Planning	Selecting actions or sequences of actions subject to objectives and constraints
State estimation	Inferring system conditions from measurements and models
Machine identity	Identifiers, credentials and related trust arrangements used for non-human computing entities
Validation	Evaluating whether a system is suitable for its intended use

The meanings above are introductory summaries; detailed definitions may differ according to the relevant discipline, standard or application.

Selected academic books

The following publications provide established technical foundations for further study.

• Siciliano, Bruno, and Oussama Khatib, editors (2016). Springer Handbook of Robotics. 2nd edition. Springer. ISBN 978-3-319-32552-1.
• Lynch, Kevin M., and Frank C. Park (2017). Modern Robotics: Mechanics, Planning, and Control. Cambridge University Press. ISBN 978-1-107-15630-2.
• Thrun, Sebastian, Wolfram Burgard, and Dieter Fox (2005). Probabilistic Robotics. MIT Press. ISBN 978-0-262-20162-9.
• LaValle, Steven M. (2006). Planning Algorithms. Cambridge University Press. ISBN 978-0-521-86205-9.
• Sutton, Richard S., and Andrew G. Barto (2018). Reinforcement Learning: An Introduction. 2nd edition. MIT Press. ISBN 978-0-262-03924-6.

Standards and institutional resources

• ISO 8373:2021 — Robotics Vocabulary — International Organization for Standardization.
• ISO 10218-1:2025 — Industrial Robot Safety — International Organization for Standardization.
• ISO 10218-2:2025 — Industrial Robot Applications and Robot Cells — International Organization for Standardization.
• Measurement Science for Robotics and Autonomous Systems — NIST.
• Perception Performance of Robotic Systems — NIST.
• NIST SP 800-82 Revision 3 — Guide to Operational Technology Security — NIST.
• NIST SP 800-207 — Zero Trust Architecture — NIST.
• NIST SP 800-57 Part 1 Revision 5 — Recommendation for Key Management — NIST.
• NIST Artificial Intelligence Risk Management Framework — NIST.

See also

• Physical AI
• Robotics
• Autonomous Systems
• Embodied AI
• Artificial Intelligence
• Computer Vision
• Sensor Fusion
• Robot Kinematics
• Robot Control
• Motion Planning
• Machine Learning
• Reinforcement Learning
• Robot Foundation Models
• Machine Identity
• Physical AI Identity
• Robot Authentication
• Robot Cybersecurity
• AI Safety
• AI Governance