Structured glossary – Embedded AI

A discipline aimed at organizing, structuring, selecting, transforming and tracing the data used by embedded AI systems, to guarantee its quality, availability, operational relevance and traceability throughout the system’s lifecycle, subject to real-time, hardware resource and safety constraints.

Raw information derived from observations of the physical or logical world, primarily produced by embedded sensors, and used as input for embedded AI processing and models for local inference and decision-making.

A coded representation, usually binary, of physical or logical information, intended to be processed by an embedded system running AI algorithms, under constrained computing, memory and latency conditions.

A structured collection of data used for training, validating or testing AI models, generally compiled prior to deployment and serving as a reference for models intended to be run on embedded systems.

An individual instance of a dataset representing a single observation (sensor measurement, image, signal, sequence), used as input to an embedded AI model during training or inference.

A measurable variable or attribute derived from a sample, describing a relevant aspect of the data and serving as an input representation for an embedded AI model, often extracted or calculated locally under resource constraints.

The set of features used to represent data in an on-device AI model, defining the dimension and structure of the model’s input, with a direct impact on the complexity, performance and cost of inference.

A separate dataset used to objectively measure the final performance of an AI model prior to its integration and deployment in an embedded system.

A value, class or category associated with a sample in the context of supervised learning, serving as a reference for the training and evaluation of embedded AI models.

The process of selecting a representative subset of data from a larger dataset, used to limit volumes, balance classes or adapt the data to the constraints of training and embedded deployment.

A subset of a dataset used to adjust the parameters of an AI model, typically processed offline prior to the model’s deployment on an embedded platform.

A subset of data used during the training phase to evaluate a model’s interim performance, adjust hyperparameters, and prevent overfitting prior to on-device deployment.

A document providing a concise and standardized description of a dataset’s characteristics (origin, composition, collection methods, intended uses, limitations, identified biases), to ensure the transparency, traceability and compliance of the embedded AI models derived from it.

A continuous or event-driven stream of data produced by on-board sensors and processed locally by AI algorithms, often subject to constraints regarding latency, bandwidth and energy consumption.

Data collected and processed within strictly defined timeframes to enable an immediate or near-immediate response from an embedded AI system, particularly in critical or autonomous applications.

The set of properties (accuracy, consistency, completeness, freshness, noise) that determine the reliability of the inferences produced by an embedded AI system and its ability to operate safely and robustly in real-world conditions.

A situation in which certain expected information is unavailable or incomplete, requiring appropriate management mechanisms to preserve the stability and safety of the embedded AI.

The ability to trace the origin, transformations, versions and uses of data within an embedded AI system, in support of safety, certification, auditability and regulatory compliance requirements.

The set of technical mechanisms designed to protect data used by embedded AI against unauthorized access, tampering and loss, whilst respecting the system’s computational and energy constraints.

The process of associating descriptive information (labels, categories, metadata) with data, generally carried out prior to embedded deployment, to enable the training and evaluation of supervised AI models intended for local inference.

A technique involving the artificial generation of new samples from existing data (geometric transformations, noise, signal variations), to improve the robustness and generalization of AI models intended for on-device execution.

An embedded hardware device that measures a physical quantity (temperature, pressure, light, sound, movement, etc.) and converts it into an electrical or digital signal that can be processed by embedded AI models, generally in real time and under resource constraints.

An optical sensor that converts light into digital signals to produce images that can be processed by on-board computer vision algorithms for perception, detection or classification.

An asynchronous optical sensor that does not capture images at a fixed frequency but generates events only when there are local variations in brightness across the pixels, enabling on-board perception with very low latency and low energy consumption.

An active sensor that emits laser pulses to measure the distance to surrounding objects and produce a three-dimensional representation of the environment, usually in the form of a point cloud, used for perception and localisation in on-board AI.

A sensor that uses radio waves to detect the presence, distance, speed and sometimes the direction of objects; it is particularly robust in adverse weather conditions (rain, fog) and is widely used for on-board perception.

A set of sensors generally comprising an accelerometer, a gyroscope and sometimes a magnetometer, used to measure linear and rotational movements, which are essential for navigation, stabilisation and multi-sensor fusion in on-board AI.

Sensors or systems used to estimate the position or trajectory of an on-board system (GNSS, UWB, local beacons), often combined with inertial sensors to improve the robustness of on-board perception and navigation.

Sensors that capture sound or vibration signals (microphones, ultrasonic sensors), used in embedded AI for speech recognition, acoustic event detection, diagnostic or monitoring.

Sensors measuring short-range distances (ultrasound, infrared, short-range ToF), used in robotics, connected devices and embedded systems for obstacle detection and interaction with the immediate environment.

Sensors measuring physical or chemical parameters of the environment (temperature, humidity, pressure, gases, particles, vibrations), used in embedded AI for monitoring, contextual analysis and anomaly detection.

Sensors measuring biological or physiological signals (heart rate, muscle or brain activity, oxygen saturation), used in embedded AI systems for healthcare, wearables and human monitoring.

Sensors measuring mechanical forces or pressures, used in embedded AI for robotics, human–machine interaction and the control of physical systems.

Optical sensors capturing information across multiple spectral bands, enabling detailed analysis of materials or the environment, used in embedded AI in specialised fields (agriculture, environment, defence).

A sensor incorporating local processing, filtering or feature extraction capabilities, enabling the reduction of transmitted data volumes and the optimisation of the performance and energy consumption of embedded AI.

A mechanism ensuring the temporal alignment of data from heterogeneous sensors, essential for sensor fusion and the consistency of inferences in embedded AI.

The association of a time reference with sensor data to ensure its consistency, traceability and correct use in real-time embedded systems.

A gradual change in a sensor’s characteristics due to the environment, wear and tear, or operating conditions, which may degrade data quality and the performance of embedded AI systems.

A set of operations applied to data from embedded sensors, after acquisition and before inference, aimed at transforming raw data into inputs that can be used by embedded AI models, whilst respecting the system’s constraints regarding latency, resources, determinism and safety.

Preprocessing carried out locally on the embedded device (microcontroller, SoC, DSP, NPU), forming an integral part of the inference chain, and aimed at improving the quality of model inputs whilst controlling computational, memory and energy costs.

Operations designed to correct, filter or ignore inconsistent, erroneous or unusable data from sensors, in order to limit inference errors and undesirable behaviour in embedded AI.

Mechanisms for handling the occasional absence of measurements (controlled ignorance, default values, simple estimation), in order to maintain stable, safe behaviour of the embedded system.

Digital processing aimed at reducing noise and disturbances in sensor signals (frequency filtering, smoothing), in order to improve the robustness of the model inputs without introducing excessive latency.

A set of software mechanisms designed to correct residual biases or drifts in digital data (offset, slow drift), in order to maintain the consistency of the model inputs over time.

Transformation of numerical values to adapt them to the range expected by embedded models and accelerators, directly linked to quantification and digital representation constraints.

Conversion of input data to reduced-precision representations (e.g. integers), in order to optimise execution on embedded targets whilst controlling the error introduced during inference.

Structural transformation of data (dimensions, channel organisation, numerical formats) to correspond exactly to the inputs expected by the embedded AI model.

Representation of discrete or categorical information in a form that can be utilised by the model (binary encoding, compact embeddings), with a focus on efficiency and embedded compatibility.

The process of transforming raw data into relevant variables that reduce the complexity of the inputs, in order to improve the robustness, interpretability and efficiency of embedded inference.

Techniques aimed at limiting the number of input variables whilst retaining useful information, in order to reduce computational load, memory usage and latency on the embedded target.

The division of a data stream into sliding or fixed time segments, enabling local analysis adapted to the real-time constraints of embedded systems.

Logical division of data (signal, image, sequence) to restrict the area or period of interest analysed by the model, thereby reducing the cost of embedded processing.

Software adjustment of the sampling frequency of digital data to adapt the volume of information to the computational capabilities and requirements of the embedded application.

Alignment of data streams from multiple sensors to ensure the temporal consistency of inputs to a multimodal embedded AI model.

Detection and processing of abnormal or inconsistent digital measurements to prevent erroneous or unstable inferences in disrupted embedded environments.

A preprocessing pipeline designed to produce reproducible results within bounded execution times, essential for integrating AI into real-time and mission-critical embedded systems.

Preprocessing optimised to minimise the use of hardware resources (computing power, memory, energy), in order to preserve the battery life and overall performance of the embedded system.

A structured and validated sequence of data processing steps between acquisition and inference, considered an integral component of the embedded AI system.

A process designed to verify that the preprocessing chain meets functional, timing and safety requirements, and that it does not degrade the overall system performance of the embedded AI under operational conditions.

Based on statistical learning and neural networks; this is an approach to AI based on models inspired by the functioning of the human brain, in particular artificial neural networks. It excels at learning from big data and processing complex signals (image, sound, text).

An approach based on explicit representations of knowledge (rules, logic, ontologies) and inference mechanisms. It prioritises structure, explainability and the manipulation of abstract concepts.

AI that combines the strengths of the two previous approaches to meet requirements for robustness, explainability and adaptability: learning capability on the one hand, and explainability and control on the other. It is particularly relevant for embedded systems requiring robustness, safety and adaptability.

Compiled languages offering fine-grained control over hardware and resources, widely used for the development of embedded AI, drivers and low-level optimisations on CPUs, GPUs and NPUs.

A modern compiled language guaranteeing memory safety and secure concurrency management, increasingly used for critical embedded applications and high-performance, reliable AI systems.

A very popular interpreted language for rapid prototyping of AI models; in embedded AI, it is used via MicroPython on resource-constrained devices or via CPython on embedded platforms for running models and high-level scripts.

An object-oriented language used on mobile and embedded platforms (Android, IoT), enabling the development of applications that incorporate AI features, with automatic memory management and a rich ecosystem of libraries.

An interpreted language primarily used in web environments and certain embedded systems with a JavaScript engine, enabling the execution of AI models on the client side and the rapid deployment of interactive AI applications.

A portable and secure binary format serving as a compilation target for languages such as C, C++ or Rust, enabling the execution of high-performance compiled code in web or embedded environments.

A tool that translates source code into native executable code for a specific hardware platform. It generally performs optimisations (topological optimisation such as layer fusion, vectorisation, pruning) to improve execution.

An open-source compilation infrastructure that provides a modular compiler and optimisation tools to generate native code on various architectures (x86, ARM, RISC-V, GPU). LLVM enables the intermediate representation of programmes, facilitating the analysis, transformation and optimisation of code for various targets, including embedded systems and AI.

An open-source automatic compilation framework designed to transform pre-trained AI models into optimised low-level code, deployable on a wide variety of hardware — from the cloud to edge devices (CPUs, GPUs, FPGAs, microcontrollers). It generates low-level code using LLVM or CUDA, whilst applying advanced optimisation techniques such as quantisation, operator fusion and pruning.

An open-source compiler infrastructure providing a generic and extensible architecture for the intermediate representation of code at multiple levels of abstraction. MLIR enables the construction of compilers and optimisation tools for machine learning, facilitating the translation of high-level models into code optimised for various hardware targets (CPUs, GPUs, neuromorphic accelerators). Its multi-level nature allows for progressive optimisations, from the semantic level of the model down to machine instructions.

An MLIR-based compiler developed by Google for TensorFlow and JAX, which transforms computational graphs into code optimised for CPUs, GPUs or TPUs by applying graph fusion and eliminating redundant operations.

A widely used compiler for ARM microcontrollers and processors, offering optimisations for speed and memory size.

A compiler developed by Xilinx, specifically designed to generate code and pipelines optimised for FPGAs and AMD accelerators.

Compiler for NVIDIA platforms such as Jetson, Orin or RTX GPUs, often used in conjunction with TensorRT or DeepStream for AI inference. Converts CUDA kernels into highly optimised executable code for GPUs. In addition to topological optimisation techniques, compilers generally rely on specific optimisation techniques

A software layer responsible for loading and executing an AI model in a given environment (cloud or embedded), potentially relying on the OS to access the hardware. It optimises resource usage (CPU, GPU, memory), latency and sometimes numerical precision (FP32, FP16, INT8), to enable the model to operate efficiently.

An open-source inference engine designed for inferencing machine learning models in ONNX format. It offers optimised (quantisation, topological optimisation) and cross-platform execution, compatible with various hardware architectures (CPU, GPU, accelerators).

An optimised inference engine, integrated into embedded frameworks such as JetPack SDK or DeepStream, enabling the optimisation (quantisation, topological optimisation, pruning) and execution of AI models (from PyTorch, TensorFlow, ONNX, etc.) on NVIDIA GPUs, particularly those designed for embedded systems (Jetson, Orin) Xavier, or RTX) in embedded or real-time environments (Jetson, Orin, Xavier, or RTX).

Basic software that manages the hardware resources of an embedded system and provides the services necessary for running embedded AI applications under real-time, memory and power constraints.

A lightweight version of ROS 2 adapted for microcontrollers, enabling the integration of embedded AI perception and decision-making functions into systems with very limited resources.

A specification defining the set of instructions, registers, addressing modes and execution model visible to the software, constituting the interface between the hardware and compilers, operating systems and AI runtimes.

A processor design philosophy based on a simple and regular instruction set, favouring pipelining, parallelism and energy efficiency, particularly suited to embedded systems.

A hardware design approach in which specifications, interfaces and sometimes hardware descriptions are open and freely accessible, allowing for the modification, redistribution and customisation of hardware architectures, particularly for AI processors and accelerators.

An open, royalty-free ISA based on RISC principles and designed in a modular fashion, allowing the addition of specialised extensions (vector processing, security, AI acceleration) to cover a wide range of processors, from microcontrollers to SoCs for embedded AI.

A family of proprietary RISC processor architectures, widely used in embedded and mobile systems, benefiting from a mature software ecosystem and extensions dedicated to vector computing and AI.

A specification for high-level tensor operators proposed by ARM and used as an intermediate format in compilation pipelines to express neural network operations independently of the hardware.

A multidimensional data structure used to represent the inputs, weights and outputs of a deep learning model. A tensor can be viewed as a generalisation of matrices to an arbitrary number of dimensions (for example, a 4D tensor for a batch of RGB images).

The in-memory organisation of a tensor’s dimensions. The two most common formats are NCHW (batch, channels, height, width) and NHWC (batch, height, width, channels). The choice of layout has a significant impact on performance, depending on the hardware architecture and the hardware library used.

A computational unit that performs a multiplication followed by an accumulation (a × b + c) in a single operation, a fundamental building block of matrix and convolutional computations in AI.

A computing method that utilises vector registers and instructions to accelerate parallel operations.

An execution paradigm in which a single instruction operates on multiple data items simultaneously.

A hardware computing unit dedicated to the efficient execution of AI models (neural networks, tensor computations), optimised to reduce latency, power consumption and memory footprint on embedded platforms.

A family of hardware accelerators designed for the execution of embedded AI models on ARM architectures, optimised for quantised neural networks and low-power inference.

Specialised hardware accelerator for the inference of embedded AI models, optimised for tensor operations and quantisation, enabling fast, energy-efficient local execution on edge devices.

A family of system-on-chip solutions integrating general-purpose compute cores and dedicated AI accelerators, demonstrating the tight integration of AI computing into low-power embedded platforms.

General-purpose processor that executes logical and arithmetic instructions.

A low-power integrated processor, suitable for simple tasks or ultra-lightweight AI models.

A parallel processor optimised for matrix calculations, used for AI training and inference.

General-purpose hardware accelerator designed to efficiently run deep learning models, not just CNNs. It optimises operations typical of neural networks: matrix multiplications, convolutions, activations, quantisation, etc.

This convolutional neural network accelerator is a hardware block dedicated specifically to convolution operations and the layers typical of CNNs. It is more specialised than an NPU, often simpler and more energy-efficient, but less versatile.

A processor specialised in digital signal processing (audio, sensors).

An AI accelerator designed by Google for large-scale tensor computations.

A processor dedicated to embedded computer vision.

A circuit inspired by the functioning of the human brain, it is based on architectures composed of hardware neurons and synapses, often implemented as spiking circuits that process information asynchronously and in a massively parallel manner. For embedded systems, these architectures offer very high energy efficiency, low latency and local data processing, making them particularly well-suited to power-constrained and real-time applications, such as perception and autonomous control.

A printed circuit board on which electronic components (CPU, GPU, memory, interfaces, etc.) are mounted, serving as a physical substrate and interconnection for embedded systems and AI modules.

A hardware-reconfigurable circuit, enabling fine-tuning of AI functions.

A chip dedicated to a specific task, highly efficient for embedded AI.

A parallel architecture comprising hundreds of cores, suitable for mission-critical applications.

An integrated circuit combining a CPU, GPU, NPU, memory and communication interfaces on a single chip, offering a compact and high-performance solution.

Internal interconnect infrastructure within an SoC enabling rapid communication between processors, caches, accelerators and memory, essential for effectively utilising the memory hierarchy and maximising data throughput.

Hardware platforms combining several types of specialised computing units, processors and accelerators within a single system. This organisation allows each task to be assigned to the most efficient computing engine, optimising the trade-off between performance, power consumption and latency. For constrained embedded systems, heterogeneous architectures thus offer good energy efficiency, a degree of execution flexibility and the ability to meet real-time requirements.

A modular integrated circuit containing a well-defined subset of functionalities (e.g. processor cores, memory, input/output).

A compact electronic board integrating a processor, memory and interfaces, designed for integration into embedded systems; used for modular embedded AI architectures.

A more comprehensive embedded module than the CoM, often including specific components (GPU, FPGA, industrial interfaces, etc.); used for demanding embedded applications (vision, robotics, etc.).

A parallel architecture designed for AI, optimised for computational graphs.

Task-adaptable processor, often FPGA-based, for customising AI acceleration

Intermediate software that ensures communication, coordination and data management between different applications, systems or layers within a computing environment. It provides common services (such as message, transaction or resource management) enabling software components to work together without direct dependencies.

Fast random access memory for the temporary processing of AI data.

Memory that retains data even without a power supply. It is the opposite of RAM (volatile).

Types of non-volatile memory (NVM) for storing AI models, firmware and logs.

Emerging, non-volatile, fast and energy-efficient memory.

Fast memories integrated into the processor to reduce latency in accessing data and instructions, organised into hierarchical levels (L1 close to the core, L3 shared).

A hardware mechanism enabling data transfer between memory and peripherals or accelerators without involving the CPU, thereby reducing latency and power consumption.

A hardware paradigm in which computation (often matrix or logical operations) is performed directly within memory cells, minimising data transfers and power consumption, and improving performance for embedded AI inference.

Interfaces enabling data transfer and communication between processors, sensors and AI modules in embedded systems.

Universal standards for connecting sensors and processors over short distances, ubiquitous in the automotive, embedded electronics and industrial sectors due to their reliability and low cost. They ensure fast, reliable data exchange suited to the real-time constraints of embedded systems. For example: I²C, SPI, UART, CAN.

Wired or hybrid communication technologies designed for industrial environments, where reliability, robustness and determinism are essential. They facilitate data exchange between PLCs, sensors, actuators and control systems, often in environments subject to severe constraints (electromagnetic noise, vibrations, high temperatures). They guarantee stable, secure and predictable communications, which are essential for industrial embedded systems and real-time applications.

Local radio technologies, suited to constrained and low-power environments, widely used in home automation, smart buildings and wearable devices. They enable local communication between sensors and gateways. They facilitate mobility and integration in constrained environments, with low energy consumption. For example: BLE, Zigbee and Thread.

Extended connectivity between devices and infrastructure, such as: Wi-Fi, LTE/4G, 5G URLLC, NB-IoT, LTE-M, LoRaWAN. 5G URLLC is essential for critical applications requiring low latency, whilst NB-IoT, LTE-M and LoRaWAN are designed for large-scale deployments of low-power connected devices. Depending on the protocol, they prioritise high bandwidth, coverage or energy efficiency, and enable edge–cloud integration.

Mesh architectures enabling resilience and extended coverage, used in distributed sensor networks and smart urban environments. Examples include 6LoWPAN and Thread mesh.

Protocols designed for the frugal and scalable transmission of sensor data, widely adopted in IoT networks.

A very lightweight protocol based on the publish/subscribe model, designed for environments with limited bandwidth and power, and widely adopted in the IoT (home automation, distributed sensors, telemetry).

A protocol inspired by HTTP, but optimised for devices with limited resources. It operates over UDP, using a request/response model. It is used in industrial automation, environmental monitoring and highly constrained sensors.

A feature-rich, message-oriented protocol with delivery guarantees and security. It is heavier than MQTT or CoAP, and therefore less suitable for highly constrained devices; it is used in enterprise environments (finance, critical systems, cloud).

A historic protocol (dating from the 1970s), simple and robust, based on a client-server model, still widely used in industry, particularly for connecting programmable logic controllers (PLCs), sensors and actuators. Its simplicity and widespread adoption make it a universal standard, but it remains limited in terms of security and data richness.

A modern, object-oriented protocol designed for interoperability and the standardisation of industrial data exchange. It is widely adopted in automation, Industry 4.0 and edge–cloud architectures, as it enables the connection of heterogeneous systems. It is extensible, secure (incorporating encryption and authentication), and supports rich, hierarchical data models.

Standardised mechanisms that streamline communication between heterogeneous layers (edge, fog, cloud) and ensure compatibility between distributed systems, equipment and services. They guarantee operational continuity in hybrid architectures and facilitate multi-technology integration. Examples: REST API, gRPC, OPC UA.

Application‑level protocol designed to standardise interactions between AI agents and external tools (services, data sources, APIs, execution environments). MCP provides a structured framework for exposing capabilities, exchanging messages and orchestrating actions, enabling seamless integration of AI agents within distributed architectures (edge - fog - cloud). Created initially by Anthropic, MCP is now an open standard and is rapidly becoming the common protocol for agentic AI, much like REST became the foundational standard for the web.

An HTTP-based communication interface using resources identified by URLs; it is universal and simple, and has established itself as the web and cloud standard for exposing and consuming services.

A modern protocol based on HTTP/2 and Protobuf, designed for remote function calls, which meets the needs of distributed architectures and microservices, with fast, efficient exchanges, including bidirectional streaming.

An approach to artificial intelligence aimed at producing new content (text, images, signals, code) based on trained models. In the context of embedded AI, generative AI is implemented in specialised and constrained forms, enabling local generation adapted to the computational, memory, latency and energy consumption limitations of embedded platforms.

An AI system designed to be adaptable to a variety of tasks. In embedded AI, these systems are generally specialised, distilled or restricted to targeted functions so that they can be executed locally and integrated into specific embedded applications.

A large-scale artificial intelligence model, trained in a self-supervised manner on vast datasets. In embedded AI, foundation models primarily serve as a starting point for compact or specialised variants, adapted for local inference on constrained systems.

The process of designing and structuring queries to guide the behaviour of a generative model. In embedded AI, prompt engineering aims to maximise the relevance, robustness and predictability of generated outputs, whilst keeping computational and memory costs under control.

A technique combining a generative model with a knowledge base search mechanism. In embedded AI, RAG relies on local, embedded or partially synchronised knowledge bases to limit dependence on remote connectivity whilst enriching the generated responses.

A compact vector representation of objects such as words, phrases, images or signals, organised to reflect their semantic or functional similarities.

A basic processing unit used by language models to represent fragments of text or symbolic data. In embedded AI, managing the number and size of tokens is a key factor in controlling latency and memory footprint.

The generation by a generative model of an incorrect, incoherent or factually unfounded response, presented as plausible. In embedded AI systems, hallucinations pose a critical risk, particularly in the fields of healthcare, transport and defence, and require mechanisms for mitigation, detection or validation.

A self-supervised learning architecture designed to predict the future evolution of a state or scene in a latent space, rather than directly generating content. JEPA is particularly well-suited to embedded AI for building models of the world that enable anticipation, planning and adaptation to dynamic environments.

Two small-scale model architectures, proposed by Sapient Intelligence and Samsung respectively, which use hierarchical looping or recursive reasoning to solve problems. These models are particularly efficient in terms of space and computational time and may be of interest for embedded systems.

Internal representations enabling an embedded AI system to model the current state and likely evolution of its environment. These models support the anticipation and reasoning capabilities essential for autonomous systems.

The ability of an AI system to initiate actions, make decisions and pursue objectives autonomously within a given environment, without continuous human supervision.

An autonomous software or hardware entity capable of perceiving its environment, reasoning on the basis of its observations, and acting to achieve a goal. In embedded AI, an agent is generally deployed on a constrained platform and acts in direct interaction with the physical world.

A set of agents interacting within a shared environment, through cooperation, coordination or negotiation. In embedded AI, these systems are used for collective robotics, intelligent sensor networks or distributed cooperative devices.

Architectures based on autonomous agents capable of perception, planning and action, promoting distributed and context-aware decision-making on embedded systems.

A paradigm in which an agent acts according to explicit goals, evaluating different courses of action and adapting its strategy in response to changes in the environment.

A continuous cycle linking environmental perception, decision-making and action, characteristic of autonomous embedded systems capable of adapting to their context in real time.

The coordination of multiple agents or AI models to perform complex tasks in a coherent manner. In embedded AI, orchestration primarily concerns distributed cooperative systems operating with limited resources and connectivity.

A form of AI integrated into hardware devices capable of perceiving, deciding and acting upon the real world. Physical AI combines sensory perception, reasoning and motor action, and is optimised to operate in real time under embedded constraints.

A faithful virtual representation of a physical system enabling the simulation, analysis and optimisation of its behaviour. In embedded AI, digital twins are used to anticipate behaviour, improve resilience and support predictive maintenance.

An approach involving the integration of environmental impact reduction from the design stage of embedded AI systems, throughout their entire lifecycle, whilst taking into account computational, energy and hardware constraints.

An approach to artificial intelligence aimed at reducing the environmental impact of AI models and systems, particularly in terms of energy consumption, carbon emissions and resource use.

An approach to the design and use of AI based on resource efficiency, the limitation of data and computation, and the alignment of functional requirements with the deployed solution, particularly suited to constrained embedded systems.

A principle aimed at limiting the amount of computation required to run a model, by prioritising lightweight architectures, efficient algorithms and appropriate inference strategies.

An approach involving the co-design of the model, algorithms and hardware architecture to reduce power consumption, computational load and memory usage, whilst ensuring functional performance.

Taking thermal constraints into account when designing models and inference pipelines, in order to prevent overheating, throttling and performance degradation.

Design aimed at maximising hardware lifespan, reparability, modularity and reusability, whilst minimising the overall environmental impact.

A strategy aimed at reducing waste and reusing components.

The effect on the environment associated with the production, use and end-of-life of a system.

An assessment of the environmental impacts of a product or system across all stages, from manufacture to end-of-life.

An assessment method that simultaneously takes into account performance, energy, memory, carbon footprint, material costs and lifespan.

A design approach guided by resource constraints (energy, memory, computation, thermal), rather than by algorithmic performance alone.

Standardised benchmarks from MLCommons for evaluating the performance of AI inference on embedded platforms. MLPerf Edge targets edge systems (SoC, CPU, GPU, NPU) and measures latency, throughput, power consumption and accuracy, whilst MLPerf Tiny is dedicated to microcontrollers and TinyML environments with severe resource constraints.

A suite of benchmarks developed by BenchCouncil designed to evaluate the performance of artificial intelligence systems for the Internet of Things (AIoT). AIoTBench offers a set of representative workloads covering various embedded AI scenarios, such as computer vision, audio processing and sensor data analysis, and enables the comparison of models, software platforms and hardware capabilities.

The ability to provide a service that can be relied upon. This implies:

The simultaneous existence of availability for authorised users only, confidentiality and integrity.

The ability of an embedded AI system to function correctly despite disruptions, uncertainties or attacks.

The degree to which a system, component or software element (code, documentation, specifications) is free from errors and correctly meets the defined requirements.

The set of properties (accuracy, consistency, completeness, freshness, noise) that determine the reliability of the inferences produced by an embedded AI system and its ability to operate safely and robustly in real-world conditions.

The origin and history of the data used to train or feed an AI system, which is essential for ensuring its reliability and compliance.

An imbalance or systematic distortion in the results produced by a model, often caused by biases present in the data or algorithms.

The principle that AI and machine learning systems must produce decisions that do not unfairly discriminate against individuals or groups, particularly on the basis of sensitive characteristics such as gender, ethnicity or age.

A set of techniques enabling the understanding and justification of decisions made by AI.

The ability to understand and explain the decisions or predictions of an artificial intelligence model.

The ability to provide a comprehensible justification for the decisions or predictions of an AI model, often required in qualification and certification processes.

The ability to track the origin, evolution and modifications of a model, software or hardware component, which is essential for qualification and certification. More specifically for embedded AI, this involves documenting the development, training and deployment stages of an AI system to ensure its transparency and auditability. Broadly speaking, it involves tracking the data, decisions and components of an AI system.

A systematic and independent examination of processes, evidence and results to confirm compliance with standards and regulations.

The ability to verify, monitor and evaluate the operation of an AI system, particularly in the event of an incident or deviation; essential for embedded systems in regulated sectors.

A detailed inventory of software components (libraries, dependencies, modules) used in an embedded system. This is used for traceability, cybersecurity and regulatory compliance.

Protection of AI models, algorithms and data used in siloed sectors where critical systems predominate. Given that critical systems rely heavily on embedded AI systems for their autonomous operations, navigation and data processing, it is essential to ensure their security. Consequently, the cybersecurity of embedded AI systems involves protecting AI-based critical systems against hacking.

AI systems to enhance the cybersecurity of critical systems in terms of security. Embedded AI systems can detect, prevent and respond to cyber threats in real time, which is crucial for the components of critical systems.

Potential security breach. Threats to cyber-physical systems include the destruction of information and/or other resources; the theft, deletion or loss of information and/or other resources; the corruption or alteration of information; the disclosure of information; and the disruption of services.

Implementation of protective measures aimed at reducing the vulnerability of assets by limiting the consequences of accidents and preventing security incidents.

An approach that integrates the protection of personal data from the earliest stages of designing an AI system. (e.g. biometric sensors).

A methodology that incorporates security mechanisms from the development phase of a system to prevent vulnerabilities.

Mechanisms built in to monitor and correct unexpected or dangerous behaviour in an AI model.

The unauthorised extraction, reverse engineering or malicious exploitation of AI/ML models deployed in critical systems. This can occur in real time during the operation or maintenance of critical systems.

A testing technique in which a team simulates attacks or malicious use to identify vulnerabilities in an AI system.

A malicious action involving the slight modification of an input (for example, by adding imperceptible noise) in order to deceive a machine learning model. These perturbations are generally invisible to humans but cause the model to make a prediction error.

The ability of an AI model to withstand adversarial attacks, i.e. to maintain acceptable performance even when its inputs have been intentionally perturbed. Robustness is a sought-after property in critical or embedded systems exposed to potential manipulation.

A defence technique involving the incorporation of misleading inputs into the training process to increase the model’s robustness. The model thus learns to recognise and neutralise malicious perturbations, making it one of the most effective methods for improving resistance to adversarial attacks.

A security mechanism ensuring that only authenticated and integrity-checked software can be executed when an embedded system boots up. In embedded AI, Secure Boot protects models, inference chains and critical components against unauthorised modifications from the moment power is applied.

A hardware-isolated environment within a processor that enables the secure execution of code and the storage of sensitive data. In embedded AI, the TEE is used to protect models, cryptographic keys and certain critical decisions against software or hardware attacks.

All activities aimed at demonstrating that a system, component or process is fit to fulfil its function in a given context. In embedded AI, this includes the validation of models and runtime environments.

The ability to assess the implementation of requirements in order to determine that they have been met (according to ARP4754A).

Verification that the system has been built correctly, in other words that it complies with the defined requirements, specifications and standards; validation consists of verifying that the right system has been built, i.e. that it effectively meets the needs and expectations of users.

Demonstration that a characteristic or property of a product meets the stated requirements.

Third-party attestation regarding products, processes, systems or individuals. This may be a written guarantee that a system or component complies with its specified requirements and is acceptable for operational use. It may also be a formal demonstration that a system or component complies with the specified requirements and is acceptable for operational use. For embedded systems, this is the process of confirming that a system or component complies with its specified requirements and is acceptable for operational use.

Documents, test results or reports that demonstrate a system’s compliance with safety, robustness or regulatory requirements. “Evidence” is essential in certification processes.

Argument used to justify certain claims regarding a system, based on evidence relating to both the system and the environment in which it operates. The claims may concern any property of the system, such as reliability or safety.

The principle that those involved in the design and use of AI must be identifiable and accountable for the decisions made by the systems. This is particularly important for autonomous embedded systems, especially in the event of an incident or dispute. Accountability is a property that ensures that an entity’s actions can be attributed solely to that entity.

The set of successive phases covering the design, development, integration, validation, conversion, optimisation, deployment, operation, evolution and end-of-life of an embedded AI system, from the initial specifications through to the system’s decommissioning.

The stage during which the functional objectives of the embedded AI, the hardware and operational constraints, as well as the performance, robustness, traceability and governance requirements necessary for its integration into an embedded system are defined.

The process of designing, training and adapting an AI model intended for execution on an embedded platform, taking into account computational, memory, latency and deterministic behaviour constraints.

The process of transforming a trained model into a format and representation compatible with the embedded device (runtime, inference library, hardware accelerator), with a view to its integration and local execution.

A set of techniques aimed at reducing the cost of running an embedded model (latency, power consumption, memory footprint), whilst minimising the impact on functional performance.

A versioned set of elements required to run an embedded AI system on the final target (converted model, parameters, processing chains, configurations), constituting the traceable and versioned operational deployment unit.

The phase during which the AI model, associated processing chains and monitoring mechanisms are integrated into the overall embedded system, interacting with the hardware, sensors and existing functions.

Commissioning of an on-device AI system on the final target, including configuration, final optimisation, verification of expected behaviour and documentation of the system’s initial state.

Phase during which the embedded AI system operates under real-world conditions, producing local inferences and decisions, under supervision and in accordance with defined governance rules.

All activities aimed at correcting, adapting or improving an embedded AI system during its operation, whilst ensuring service continuity, the traceability of changes and compliance with the original requirements.

The phase during which an embedded AI system is taken out of service, including the controlled deactivation of AI functions, the archiving of elements necessary for traceability, and the management of final model versions.

A process aimed at demonstrating that the embedded AI system meets the defined functional and operational requirements, under conditions representative of its actual use on the embedded platform.

All activities designed to ensure that the embedded AI system is correctly implemented in accordance with the defined specifications, particularly in terms of behaviour, performance and time constraints.

Tests carried out on the embedded target or in a representative environment to assess the behaviour of embedded AI under conditions close to actual operation.

The process of verifying, during operation or following an update or recalibration, that the behaviour of the embedded AI remains compliant with the initial requirements, despite changes in usage conditions, input data or the operational context.

The set of rules, processes and responsibilities enabling the management of the design, deployment, operation and evolution of embedded AI systems, ensuring the consistency of technical and organisational decisions throughout the lifecycle.

The principle that those involved in the design and use of AI must be identifiable and accountable for the decisions made by the systems. This is particularly important for autonomous embedded systems, especially in the event of an incident or dispute. Accountability is a property that ensures that an entity’s actions can be attributed solely to that entity.

Definition of responsibilities relating to the day-to-day operation, monitoring and corrective actions on an embedded AI system, particularly in the event of a performance degradation or incident.

The ability to trace the history of decisions, model versions, parameters, processing chains and configurations throughout the lifecycle of an embedded AI system, in order to enable the auditing and justification of the choices made.

A process for managing changes to models, parameters, processing chains and configurations deployed on embedded systems, ensuring clear identification of the versions in operation.

The set of operational practices and processes dedicated to AI models deployed on embedded systems, covering deployment, monitoring, maintenance, updates and version management, subject to resource and connectivity constraints.

Mechanisms for monitoring the operation of embedded AI at runtime, to detect abnormal behaviour, performance deviations or unforeseen situations.

Continuous or periodic observation of the behaviour and performance of embedded AI during actual operation, whilst adhering to the system’s computational and energy constraints.

Collection and reporting, on the on-device target, of metrics relating to AI operation (inference time, load, input quality, confidence scores, degradation events), to enable analysis and supervision compatible with on-device constraints.

Gradual degradation of a model’s performance in operation, linked to changes in input data, usage context or environment, which may require recalibration, updating or revalidation of the on-device system.

The set of mechanisms designed to detect, characterise and address performance drifts in an embedded AI system, including thresholds, alerts, degraded modes, recalibration and versioned updates.

Planned or triggered adjustment of an embedded AI system (parameters, thresholds, associated calibrations) to maintain compliant performance during operation, whilst ensuring the necessary traceability and revalidation.

The activity of monitoring the behaviour of an embedded AI system over time after it has been commissioned, to confirm that expected performance is being maintained and to identify any need for adjustment or updating.

The ability of an embedded AI system to adopt safe, predictable and documented behaviour in the event of performance loss, degraded data or non-nominal operating conditions.

A mechanism for remotely deploying a new version of a model, setting or configuration to an embedded target, ensuring version identification and service continuity.

The ability to restore a previous version of a model or configuration on an embedded target, particularly in the event of a performance regression, instability or an incident detected during operation.

Regulatory framework proposed by the European Union to govern the use of AI according to its level of risk, with specific obligations for critical systems.

European regulation harmonising the rules on the processing of personal data in the EU, which came into force on 25 May 2018. It strengthens individuals’ rights and makes data controllers accountable.

Having entered into force in January 2024, the EU Data Act (Regulation (EU) 2023/2854) has been fully applicable since September 2025 following the expiry of its transitional period. This legislation establishes a harmonised framework setting out new rules on the access, use and sharing of industrial data generated by connected products and services in the EU, across all economic sectors.

The European Chips Act (Regulation (EU) 2023/1781) aims to strengthen the EU’s semiconductor capacity, support research and development and develop advanced production capabilities. Its objective is to achieve 20 per cent of global chip production by 2030. It also provides for a coordination mechanism to secure the supply chain in the event of a crisis. It promotes the emergence of a European ecosystem capable of designing and producing components optimised for embedded AI and cyber-physical systems.

The Data Governance Act (Regulation (EU) 2022/868) sets out the governance and mechanisms for data sharing, in particular the reuse of certain public sector data, the framework for data intermediaries and “altruistic” data sharing.

The new European Product Liability Directive (2024/2853) was adopted on 23 October 2024 and must be transposed into the national laws of Member States by 9 December 2026. It modernises liability rules to include digital products, AI and circular business models.

A European regulation adopted in 2019, which establishes the European cybersecurity certification framework for ICT products. It gave rise to future certification schemes such as EUCC, EUCS, EU5G, etc. It does not create direct obligations for manufacturers, but it structures the trust framework and certification schemes. Its aim was to harmonise cybersecurity certification across Europe.

A European regulation that came into force on 10 December 2024, aimed at improving the cybersecurity of products containing digital elements (PDEs) and securing the Internet of Things. It imposes legal cybersecurity obligations on all products with digital elements (PDEs), including connected devices, embedded software, industrial systems, etc. It aims to ensure that all digital products sold in Europe meet a minimum level of cybersecurity throughout their lifecycle.

Regulation (EU) 2023/988 on general product safety, applicable from December 2024. It requires manufacturers to ensure the safety of consumer products, including those incorporating software or AI, and strengthens market surveillance obligations.

A measure introduced by the EU’s Circular Economy Package, aimed at providing a digital passport containing essential information about a product (composition, reparability, durability, traceability). It is being gradually applied to several product categories, including electronic equipment.

Directive (EU) 2022/2555 strengthening cybersecurity requirements for essential and important sectors. It imposes increased obligations regarding risk management, system security and incident reporting, directly impacting embedded systems and critical infrastructure.

Directive 2014/53/EU on radio equipment now incorporates strengthened requirements regarding cybersecurity, data protection and fraud prevention for connected devices. The associated delegated acts impose specific obligations on manufacturers of radio devices incorporating software or AI functions.

European cybersecurity certification scheme for ICT products, developed under the Cybersecurity Act. It provides a harmonised framework for certifying the security of hardware and software components, relevant to embedded systems and products incorporating AI.

Quality management for AI systems

Risk management for AI systems.

Logging, transparency and human oversight (prEN 18229-1): managing three aspects of trust in AI systems: traceability, user information and oversight by human operators.

Accuracy and robustness (prEN 18229-2): management of two aspects of trust in AI systems: accuracy and robustness.

Standard on the quality and governance of datasets used in AI systems.

Management (and reduction) of bias in data, in order to meet fairness requirements.

Specifications on the cybersecurity of AI systems, analysis of threats and ways to respond to them.

Assessment of conformity with all standards, carried out by authorised bodies.

The first international standard dedicated to artificial intelligence (AI) management systems, published in December 2023, which provides a framework for the development, use and governance of AI systems to ensure responsible, ethical and secure use.

Published on 20 December 2023, it provides a standardised framework for the development and management of AI systems, covering all aspects of the lifecycle of software incorporating AI.

Provides a comprehensive framework for identifying, assessing and managing risks associated with artificial intelligence, drawing on the principles of ISO 31000 to ensure effective integration into organisations’ activities.

A coherent set of functions provided by an embedded system, some of which are based on AI models running on the target platform. An embedded AI application combines sensors, processing and decision-making logic to produce a behaviour, capability or operational service integrated into the system.

A structured description of an operational scenario in which one or more embedded AI functions contribute to achieving an identified operational objective. A use case specifies the context, sensor inputs, expected outputs and associated execution constraints.

A software function deployed on an embedded system, implementing a processing chain and one or more AI models to produce an actionable output (information, alert, estimate, decision or command) in interaction with the environment.

A description of a real-world usage context characterised by an environment, variable operating conditions, time constraints, and nominal or degraded situations to which the embedded AI must respond reliably.

Advantage gained by running AI models directly on the embedded target, enabling compliance with requirements for responsiveness, autonomy, service continuity, resilience or control of data flows.

Applications in which embedded AI analyses data from industrial sensors to carry out inspections, monitor processes and detect anomalies directly on the equipment. Embedded processing enables immediate responsiveness, robustness in industrial environments and a reduction in data flows to central systems.

Applications where embedded AI supports perception, navigation, manipulation and interaction between robots and operators. Embedded processing ensures latency compatible with movement and physical interaction, as well as autonomous operation in dynamic environments.

Applications incorporating embedded AI to interpret the environment, assist with driving, manage trajectories and support autonomous functions. Embedded systems enable fast and reliable local decision-making, independent of network connectivity.

On-board applications in aircraft, drones or space systems using AI for perception, surveillance and decision support. On-board AI provides autonomy, resilience and service continuity in isolated or highly constrained environments.

Applications in which embedded AI contributes to surveillance, situational analysis and decision support. Embedded processing enables robust, offline and controlled operation, which is essential in critical contexts.

Applications integrating embedded AI into medical devices, diagnostic equipment or monitoring systems. Local processing promotes responsiveness, operational continuity and better control over sensitive data.

Applications utilising embedded AI to monitor, diagnose and optimise energy equipment or distributed infrastructure. Embedded processing enables efficient local management and large-scale operation of autonomous devices.

Applications in which embedded AI is used to optimise the use of network resources and equipment, detect anomalies and improve service quality on edge computing. Embedded systems enable rapid adjustments as close as possible to the traffic and equipment.

Applications using embedded AI for energy management, equipment monitoring and the security of buildings or structures. Embedded processing facilitates local decisions tailored to actual usage and varying conditions.

Applications utilising sensors and on-board vision to monitor the environment, analyse crops or detect natural events. On-board AI enables autonomous, energy-efficient devices capable of operating for long periods without continuous connectivity.

Applications incorporating compact AI models for interaction, personalisation and activity recognition. On-device processing ensures a responsive experience, low energy consumption and independence from remote infrastructure.

Applications in which embedded AI supports audiovisual analysis, real-time interaction, augmented reality or immersive experiences. Embedded processing enables fluid and immediate interactions on autonomous devices.

The ability of an on-board system to interpret its physical environment using data from sensors (image, radar, LiDAR, audio, inertial, biometric, etc.), to produce a representation that can be used for decision-making or to control a physical system.

The function of identifying the presence or occurrence of objects, events or situations of interest in input data, for example the detection of a fault, an obstacle, an incident or a change of state, with an output that is generally binary or localised.

A function aimed at associating a sensor observation with a predefined category, such as object type, equipment status, environmental condition or activity, to guide a decision or local action.

A function enabling data, particularly an image or a signal, to be divided into homogeneous regions or segments corresponding to areas of interest, in order to precisely locate objects, defects or phenomena relevant to the embedded application.

A function that produces a continuous value representing a state or a physical quantity, such as a position, velocity, orientation, level or intensity, based on sensor measurements, to feed into a control or supervision logic.

A function consisting of anticipating the future evolution of a state, behaviour or observed phenomenon, for example equipment wear, a probable trajectory or a performance drift, to enable proactive action by the embedded system.

A function designed to identify significant deviations from a learned or defined nominal operating state, to detect faults, incidents or unusual situations early on in real-world environments.

A function enabling the identification of an activity, gesture or behaviour based on time series or multimodal signals, such as data from inertial, biometric or acoustic sensors, particularly in wearables and interactive systems.

A function enabling direct and responsive interaction between a user and an embedded system, via voice, gestures, context or sensory interfaces, with local processing ensuring low latency and service continuity.

A functional organisation in which information derived from perception feeds into a local decision, which then triggers an action or command on the system or its environment, forming a closed loop executed under real-time constraints.

The principle whereby AI functions are executed directly on the embedded platform, without systematic reliance on a remote infrastructure, to guarantee responsiveness, operational autonomy and service continuity in constrained environments.

A requirement stipulating that the results produced by AI must be available within a timeframe compatible with the application, for example to trigger an action, assist an operator or control a physical system, with limited and controlled response times.

The ability of an AI-integrated embedded system to operate sustainably within a limited energy budget, whether powered by a battery, energy recovery, or subject to strict consumption constraints.

The ability of the embedded AI system to maintain useful and predictable behaviour despite the variability of real-world conditions, sensor noise, environmental disturbances or non-nominal situations.

A context in which the embedded system cannot rely on continuous or high-speed network access, requiring autonomous decision-making and processing capabilities directly on the target.

Deployment of the same embedded AI function across a large number of devices, requiring consistent behaviour, version control and industrial-scale operation of the embedded fleet.

A use case in which an embedded system analyses images or video streams from cameras to detect defects, assess compliance or sort objects, with local execution ensuring responsiveness and robustness in industrial or operational environments.

A use case where signals from sensors (vibration, acoustic, current, temperature, pressure) are analysed by on-board AI models to anticipate failures, identify operational deviations and plan appropriate maintenance actions.

Use cases involving the timely detection of rare or abnormal events, such as an incident, dangerous behaviour or a threshold being exceeded, to trigger an alert or immediate action by the embedded system.

Use case in which multi-sensor perception and local decision-making support driving assistance, navigation or partial autonomy functions, with high real-time and reliability requirements.

Use case where an on-board system dynamically adjusts energy management strategies (charging, equipment control, prioritisation) based on local measurements and estimates, in order to improve energy efficiency without continuous reliance on a remote infrastructure.

Use cases in which embedded AI provides recommendations, alerts or diagnostics directly in the field, in contexts with limited or no connectivity, to support human or automated action.

Use cases where a wearable device locally analyses data from inertial, biometric or physiological sensors to identify activities, gestures or states, whilst meeting strict energy efficiency requirements.

Use cases in which embedded devices deployed on infrastructure carry out continuous or periodic diagnostics of the condition of equipment or structures, producing actionable indicators even under conditions of intermittent connectivity.

The limited ability of an on-device AI system to adjust certain parameters or thresholds during operation to maintain its performance in the face of changing environments, subject to resource and safety constraints.

An organisation in which several embedded systems cooperate to produce a collective perception or decision, whilst retaining predominantly local execution.

Applications in which embedded AI is used as a lever for digital frugality, minimising the energy footprint, the volumes of data processed and dependence on remote infrastructure, whilst maintaining operational utility.

Applications utilising embedded AI to observe, analyse and act locally on environmental phenomena, using autonomous, distributed and resource-efficient devices.

Applications in which embedded AI helps to build trust among users and stakeholders through controlled local decision-making, increased operational transparency and integration that respects human and social contexts.