Synthetic Data, AI & Robotics Glossary
The definitive reference for physical AI terms across simulation, robotics, computer vision, aerial, and maritime autonomy, defined by the team building the infrastructure behind it.
A
3D Asset Library
A 3D asset library is a catalog of digital models — vehicles, vessels, terrain, buildings, people — used to populate simulated environments. The richer and more realistic the library, the more varied and useful the synthetic data generated from it.
Action Segmentation
Action segmentation is the task of dividing a continuous video sequence into time-bound segments that each correspond to a distinct action, like "approaching dock" or "pedestrian crossing." It's foundational for AI systems that need to understand not just what's in a scene, but what's happening in it over time.
Active Learning
Active learning is a training approach where a model identifies the data points it's least confident about and flags those for labeling, instead of learning from a randomly labeled set. It helps teams get more value from every label by focusing effort where the model actually needs it.
Aerial Autonomy
Aerial autonomy is a drone's or aircraft's ability to navigate, make decisions, and complete missions without constant human piloting. Building it reliably means testing against thousands of edge cases, weather, terrain, traffic, that are too dangerous or expensive to fly in the real world.
Aerial ISR (Intelligence, Surveillance, Reconnaissance)
Aerial ISR describes drone or aircraft missions flown to gather information — imagery, video, signals — about an area or target. ISR systems lean heavily on perception AI, which makes them only as good as the diversity of conditions they've been trained and evaluated on.
Airspace Deconfliction
Airspace deconfliction is the process of detecting and resolving potential conflicts between multiple aircraft or drones sharing the same airspace, keeping flight paths safely separated. As drone traffic grows, doing this autonomously is becoming a requirement.
Altitude Hold
Altitude hold is a flight control mode that automatically maintains an aircraft's current altitude without pilot input, freeing up attention for navigation or mission tasks. It's a basic building block of more advanced autonomous flight behaviors.
Anomaly Detection
Anomaly detection identifies data points, events, or behaviors that deviate from expected patterns, like a vessel moving erratically or a sensor reading that doesn't make sense. It's especially valuable in physical AI, where rare anomalies often represent the highest-stakes failures.
Automated Labeling
Automated labeling uses AI models — often trained on synthetic data — to generate annotations for real-world imagery without manual review. It removes one of the slowest, most expensive bottlenecks in building a training dataset.
Autonomous Vehicle (AV)
An autonomous vehicle is any vehicle — ground, air, sea, or space — capable of sensing its environment and operating with little or no human intervention. Every AV depends on perception, planning, and control systems that have to be tested far beyond what real-world driving or flying can practically cover.
B
Balanced Data
Balanced data is a dataset in which the categories or scenarios a model needs to learn are represented in roughly equal proportion, instead of being dominated by the most common cases. Synthetic data generation is one of the fastest ways to balance a dataset without waiting to collect more real-world examples.
Behavior Cloning
Behavior cloning trains a robot or autonomous system to imitate actions recorded from a human or expert demonstration. It's a fast way to bootstrap a policy, though it's only as good as the diversity of demonstrations it learns from.
Benchmark
A benchmark is a standardized test or dataset used to measure and compare AI model performance under consistent conditions. Good benchmarks have to reflect real operating conditions — not just the easy cases.
Beyond Visual Line of Sight (BVLOS)
BVLOS refers to drone operations conducted beyond the pilot's direct line of sight, relying on sensors, automation, and communications links instead. It unlocks much longer-range missions, but raises the bar for how rigorously the autonomy stack needs to be tested.
Bias
Bias describes systematic errors that occur when training data underrepresents certain conditions, objects, or scenarios, causing a model to perform worse on them. Addressing bias starts with finding the gaps in data before they show up as failures in deployment.
Boundary Conditions
Boundary conditions are the extreme limits of an operating environment — the harshest weather, the tightest space, the fastest closing speed — that an AI system might face. Testing at the boundaries, not just the average case, is what separates a system that works in demos from one that works in the field.
Bounding Box
A bounding box is a rectangular annotation drawn around an object of interest in an image, marking its location for an object detection model to learn from. It's one of the most common, foundational labels in computer vision.
Burst Simulation
Burst simulation models short, intense events — a sudden squall, a flock of birds, a debris field — that occur briefly but can meaningfully disrupt a perception system. Simulating these bursts lets teams test for rare, high-impact moments without waiting for them to happen in the real world.
C
Class Balancing
Class balancing adjusts a dataset so underrepresented categories get equal weight to common ones, typically by generating more examples of the rare classes. It's a direct fix for one of the most common causes of poor model performance: not enough examples of what matters most.
Class Imbalance
Class imbalance occurs when some categories in a dataset vastly outnumber others — thousands of images of cars but only a handful of kayaks. Left unaddressed, it causes models to perform well on common cases and poorly on the rare ones that often matter most.
Collision Avoidance
Collision avoidance lets an autonomous system detect an obstacle in its path and take action to avoid it, whether that's a warehouse robot or a vessel at sea. It depends on a perception and planning stack tested against every kind of obstacle, not just the obvious ones.
Collision Imminent Alert
A collision imminent alert is a warning generated when a maritime autonomy system detects a closing trajectory with another vessel or object, with insufficient time or space to avoid impact. Generating these alerts reliably requires training and testing on a wide range of vessel types, speeds, and approach angles.
Computer Vision
Computer vision is the field of AI focused on enabling machines to interpret visual information — images, video, LiDAR, radar. It's the perception backbone of nearly every autonomous system, which means its training data has to represent the real world in all its messy variety.
Continuous Integration for AI (CI)
CI for AI applies the software practice of testing every change before it ships to machine learning models, running them against benchmark scenarios with every update. It catches regressions before they reach the field, not after.
Convolutional Neural Network (CNN)
A CNN is a deep learning architecture designed to process grid-like data such as images, using layers of filters to detect patterns like edges, textures, and shapes. CNNs remain a core building block behind most modern object detection and classification systems.
Corner Cases
Corner cases are rare, unusual scenarios at the edge of what a system was designed to handle — a pedestrian in an odd costume, a vessel doing something unexpected. They're disproportionately responsible for AI failures because they're disproportionately absent from training data.
Counter-Drone Detection
Counter-drone detection identifies unauthorized or hostile drones in a protected airspace, typically using radar, RF, acoustic, and visual sensors together. Reliable detection across drone sizes, speeds, and behaviors is hard to validate without simulating a wide range of drone threats.
Counter-UAS (C-UAS)
Counter-UAS covers the full set of technologies and tactics used to detect, track, and neutralize unmanned aircraft that pose a threat. Proving out C-UAS systems requires testing against realistic drone behaviors at a scale live flight tests can't match.
Coverage Metrics
Coverage metrics quantify how much of an AI system's operational domain has actually been tested, across weather, lighting, object types, and scenario complexity.
D
Data Annotation
Data annotation is the process of labeling raw data — drawing boxes, tracing masks, tagging classes — so a model can learn from it. Synthetic data comes with perfect, automatic annotations built in, skipping the most labor-intensive step in building a dataset.
Data Augmentation
Data augmentation expands a training dataset by creating modified versions of existing data — rotated, cropped, recolored — to improve a model's robustness. Synthetic data generation goes further, creating entirely new, realistic scenarios rather than variations of what already exists.
Data Drift
Data drift occurs when the statistical properties of incoming real-world data shift over time, causing a model trained on older data to become less accurate. Catching drift early means continuously testing against fresh data, not assuming yesterday's performance holds today.
Data Engine
A data engine is an end-to-end system that generates, labels, and delivers the exact training and test data an AI team needs, on demand. It replaces the slow cycle of collecting and manually labeling real-world data with a faster, more controllable pipeline.
Data Generation
Data generation is the process of creating new training or testing data, whether by capturing it in the real world or simulating it synthetically. Simulated generation lets teams produce exactly the scenarios they need, at a scale and speed real-world collection can't match.
Data Patch
A data patch is a targeted batch of synthetic data created specifically to fix a known weakness in a model — a missed object type, a failure in fog, a gap in a particular approach angle. Rather than collecting more general data and hoping it helps, a data patch goes straight at the problem.
Data Pipeline
A data pipeline is the sequence of steps that moves data from collection or generation through processing, labeling, and into model training, ideally with as little manual intervention as possible. A well-built pipeline is what turns new data into a better model in days, not months.
Data Privacy
Data privacy covers the practices that keep sensitive or personally identifiable information from being exposed during data collection, storage, or use in AI training. Synthetic data is one way teams reduce privacy risk, replacing or supplementing real-world data that would otherwise need protection.
Data Quality
Data quality describes how accurate, complete, consistent and representative a dataset is for the task a model needs to learn. It matters more than raw volume — a smaller, well-targeted dataset often outperforms a much larger but noisy one.
Data Scarcity
The lack of sufficient real-world training data to build or improve an AI model. In physical AI, this is common because collecting data across dangerous, rare, or hard-to-reach conditions is costly and slow. Gaps in data lead directly to gaps in model performance.
Data-Centric AI
An approach to AI development that prioritizes improving training data over tweaking model architecture. Rather than assuming the dataset is fixed, teams systematically identify what data is missing or wrong and fix it — resulting in faster, more reliable model improvements.
Depth Estimation
Depth estimation determines how far away each point in a scene is from the camera, producing a sense of 3D structure from 2D images or sensor data. It's essential for any system that needs to navigate around or interact with objects in physical space.
Digital Elevation Model (DEM)
A Digital Elevation Model is a 3D representation of a terrain's surface elevation, typically generated from aerial or satellite data. DEMs are a key input for simulating realistic terrain in geospatial and aerial training environments.
Digital Surface Model (DSM)
A Digital Surface Model captures the elevation of everything on the ground — buildings, trees, vehicles — not just bare terrain like a DEM. DSMs let simulated environments reflect what a sensor would actually see, obstacles included.
Domain Adaptation
Domain adaptation is the set of techniques used to help a model trained in one domain — say simulation — perform well in a different but related domain, like the real world. It's the technical core of making synthetic data actually useful for real-world deployment.
Domain Gap
The domain gap is the difference between simulated and real-world data that can cause a model trained on one to underperform on the other. Closing it through photorealistic rendering, sensor simulation, and domain randomization is the central challenge synthetic data exists to solve.
Domain Randomization
Domain randomization deliberately varies elements of a simulated scene — lighting, textures, object positions, weather — across many generated samples. Exposing a model to enough variation in simulation helps it learn features that generalize to the real world instead of overfitting to one look.
Downstream Model
A downstream model is the AI model ultimately trained and deployed using a given dataset, as opposed to any models used to generate or process that data along the way.
Drone Swarm
A drone swarm is a group of drones coordinating movement and decision-making to complete a shared mission. Testing swarm behavior safely at scale is virtually impossible in the real world, which makes simulation the primary tool for developing and validating swarm logic.
Drone-in-a-Box
Drone-in-a-box describes a system where a drone is deployed, charged, and stored autonomously in a fixed ground station, enabling repeated missions with minimal human involvement. Making these systems reliable means validating the full mission loop — takeoff, flight, landing, charging — across countless conditions.
Dynamic And Event-Based Classifications
This refers to labeling schemes that classify scenes based on actions and events unfolding over time — a vessel docking, a drone evading — rather than static objects in a single frame. It's increasingly important as perception systems are asked to understand behavior, not just identify things.
Dynamics Randomization
The practice of randomly varying simulation parameters such as lighting, textures, physics, and sensor noise during training so that an AI model learns to handle real-world variability. Models trained with domain randomization generalize better because they've never seen the same conditions twice.
E
Edge Case
An edge case is a scenario at the boundary of what a system was designed to handle unusual, rare, but still plausible. Most AI failures in deployment trace back to edge cases that simply weren't in the training or test data.
Edge Simulation
Edge simulation focuses on generating and testing the rare, boundary-pushing scenarios that real-world data collection rarely captures. It's how teams build confidence in a system's behavior before it ever encounters those situations live.
Electro-Optical/Infrared (EO/IR) Sensor
An EO/IR sensor combines a standard electro-optical camera with an infrared camera in a single payload, giving an aircraft visibility in both daylight and low-light or thermal conditions. Training perception models for EO/IR systems means simulating both spectrums realistically, not just the visible one.
Embodied AI
Embodied AI refers to AI systems that learn and act through a physical body — a robot, drone, or vehicle — rather than existing purely in software. Because embodied agents deal with the unpredictability of the physical world, they need training and testing environments that capture that unpredictability realistically.
Environment Randomization
Environment randomization varies the broader scene context — terrain, structures, vegetation, layout — across simulated training data, rather than just object-level details. It helps models generalize across the wide range of real-world settings they'll actually be deployed into.
Evaluation Pipeline
An evaluation pipeline is the automated system that runs an AI model against a defined set of test scenarios and metrics to measure its performance. A strong one turns testing from a one-time event into a continuous, repeatable process.
eVTOL (Electric Vertical Take-Off and Landing)
An eVTOL is an aircraft that takes off and lands vertically using electric propulsion, built for urban air mobility and short-range flight without a runway. As a relatively new aircraft category, eVTOL programs often lack the flight history other aircraft rely on, making simulation critical to building out their safety case.
F
Failure Mode
A failure mode is a specific way an AI system can go wrong, misclassifying an object, losing track of a target, misjudging distance. Cataloging failure modes systematically is the first step to fixing them, rather than discovering them in the field.
Fairness
Fairness in AI means a model performs consistently well across different groups, conditions, or populations, rather than favoring the cases best represented in its training data. In physical AI, that often shows up as equally reliable performance across object types and environments not just demographic groups.
Feature
A feature is an individual measurable property a model uses as input, like an object's size, color, or velocity. Good feature representation is what lets a model tell meaningfully different things apart.
Fidelity
Fidelity describes how closely simulated data visually, physically and statistically matches the real-world conditions it's meant to represent. Higher fidelity simulation generally means a smaller domain gap and a model that transfers more reliably to deployment.
Fine-tuning
Fine-tuning takes a pre-trained model and further trains it on a smaller, more specific dataset to adapt it to a particular task or environment. It's a fast, efficient way to specialize a general model without training from scratch.
First-Person View (FPV)
First-Person View means piloting or operating a drone using a live video feed from the aircraft's own camera, as if seeing through its eyes. FPV puts extreme demands on low-latency, robust perception, since situational awareness depends entirely on what the sensor sees.
Flight Path Planning
Flight path planning calculates a safe, efficient route for an aircraft, accounting for terrain, obstacles, airspace restrictions and mission objectives. Validating a planner means testing it against far more terrain and obstacle combinations than any single flight could cover.
Fog Simulation
Fog simulation models the visual and sensor effects of fog — reduced visibility, light scatter, sensor noise — within a synthetic environment. It lets teams test perception performance in degraded visibility without waiting for the weather to cooperate.
Foliage Penetration
Foliage penetration is a sensor's ability — often LiDAR, SAR, or specialized EO/IR — to see through tree canopy and detect objects or terrain hidden beneath it. It's a critical capability for ISR and search missions in forested terrain.
Foundation Model
A foundation model is a large model trained on broad, diverse data that can be adapted to many downstream tasks through fine-tuning, rather than built for one narrow purpose. As foundation models extend into physical AI, rigorously evaluating them across real-world operating conditions matters more than ever.
Frame-by-Frame Labeling
Frame-by-frame labeling annotates every individual frame in a video sequence, rather than sampling a subset, to preserve continuity for tasks like tracking. It's labor-intensive by hand, but synthetic data generates it automatically and perfectly, for every frame.
Frontier Scenario
A frontier scenario is a test case that pushes beyond a system's currently validated operating envelope, probing where its capabilities actually break down. Frontier scenarios are how teams find the edge of what their AI can handle before the real world finds it for them.
Full-Motion Video (FMV)
Full-Motion Video refers to continuous, real-time video feeds — typically from drones or aircraft used for surveillance — as opposed to still imagery. Training perception models for FMV analysis means working with the temporal, often noisy nature of live video, not just clean still frames.
Full-Stack Simulation
Full-stack simulation models an entire autonomous system end-to-end — sensors, perception, planning, and control within a single simulated environment, rather than testing each layer in isolation. It's the closest thing to a real-world test that doesn't require real-world hardware.
G
GAN
A Generative Adversarial Network pairs two competing models, a generator that creates synthetic data and a discriminator that tries to tell it apart from real data. The competition between them pushes the generator to produce increasingly realistic outputs.
Gaussian Splatting (3DGS)
Gaussian Splatting represents and renders 3D scenes using millions of small, semi-transparent "splats" instead of traditional meshes, enabling fast, photorealistic real-time rendering. It's increasingly used to reconstruct real-world environments for use in simulation.
Generative Synthetic Data
Generative synthetic data is created by generative AI models, rather than rule-based or procedural systems to produce new, realistic samples based on patterns learned from existing data. It's one of several approaches to synthetic data generation, often combined with simulation for physical AI.
Geofencing
Geofencing establishes a virtual perimeter around a defined geographic area that an autonomous system can be programmed to stay within or avoid. It's a foundational safety mechanism for drones and other autonomous vehicles operating near restricted areas.
Geospatial AI
The application of AI techniques to analyze and interpret location-based data, such as satellite imagery, aerial photos, and maps. Geospatial AI is used for tasks like land use classification, change detection, and infrastructure monitoring across large geographic areas.
Gimbal Stabilization
Gimbal stabilization uses a motorized mount to keep a camera or sensor steady and correctly oriented despite the motion of the aircraft carrying it. Stable footage is a prerequisite for reliable perception, since shaky imagery degrades almost every downstream model.
Ground Sample Distance (GSD)
Ground Sample Distance is the real-world size of a single pixel in an aerial image, determining how much detail can actually be resolved. Lower GSD means higher resolution which matters directly for what an object detection model can and can't reliably see.
Ground Truth Labels
Ground truth labels are the verified, correct annotations for a dataset — the answer key a model is trained and evaluated against. Synthetic data generates ground truth automatically and with perfect accuracy, since the simulation already knows exactly what's in every scene.
H
Hardware-in-the-Loop (HIL)
Hardware-in-the-loop testing connects real physical hardware — a flight controller, a sensor, an actuator — to a simulated environment, so it responds to simulated inputs as if they were real. It bridges the gap between pure software simulation and full real-world testing.
Hazard Scenario
A hazard scenario is a test case specifically designed to evaluate how an AI system responds to a dangerous or safety-critical situation. Simulating hazard scenarios lets teams validate safe behavior without ever putting a real system, or real people, at actual risk.
High-Performance Computing (HPC)
HPC refers to the large-scale computational infrastructure used to run demanding workloads like model training and large-scale simulation, far beyond what a standard workstation can handle. Generating and processing data at the scale modern AI needs almost always depends on HPC somewhere in the pipeline.
Human-in-the-Loop Evaluation
Human-in-the-loop evaluation incorporates human judgment at key points in testing, reviewing model outputs that automated metrics alone might miss or misjudge. It's especially valuable for catching subtle failures that don't fit neatly into a predefined metric.
Humanoid Robot
A humanoid robot has a human-like body plan — head, torso, arms, legs — typically to operate in environments and use tools built for people. Training humanoid locomotion and manipulation policies relies heavily on simulation, since real-world trial and error on physical hardware is slow and risky.
Hyperspectral Imaging
Hyperspectral imaging captures data across many narrow, contiguous spectral bands far more than the three in a standard color image, revealing material properties invisible to the naked eye. It's used for tasks like vegetation health analysis and material classification and requires specialized simulation to generate realistic training data.
I
Image Annotation
Image annotation labels visual data with what a model needs to learn from it — bounding boxes, masks, keypoints, class tags. Synthetic data eliminates manual annotation entirely, since every generated image already comes with exact, built-in labels.
Imbalanced Dataset
A dataset in which some classes, objects, or scenarios are significantly overrepresented compared to others. Imbalanced datasets cause models to perform well on common cases but poorly on rare ones — a frequent issue in physical AI, where critical edge cases are naturally underrepresented in real-world data.
Imitation Learning
A training approach where an AI model learns to perform a task by observing and replicating expert demonstrations, rather than learning through trial and error. Common in robotics, where a model learns control policies by mimicking human-operated examples.
Incident Replay
Incident replay recreates a real-world failure or near-miss inside a simulated environment, so teams can analyze exactly what went wrong and test fixes against the same conditions. It turns every real incident into a permanent, repeatable test case.
Industrial Robotics
Industrial robotics covers automated systems used in manufacturing, warehousing, and logistics to handle tasks like assembly, picking, and material handling. As these robots take on more autonomous decision-making, validating their perception and planning in simulation becomes essential to safe deployment alongside people.
Infrared Imaging
Infrared imaging captures the heat signature emitted by objects rather than visible light, allowing detection in darkness or through obscurants like smoke. Reliable infrared perception models need infrared-specific training data, since thermal signatures behave very differently from standard imagery.
Instance Segmentation
A computer vision technique that identifies and outlines each individual object in an image at the pixel level, distinguishing between separate instances of the same class — every individual person in a crowd, for example. It provides far more precise spatial understanding than a bounding box alone.
Intersection over Union (IoU)
Intersection over Union (IoU) is a metric measuring how well a predicted bounding box or segmentation mask overlaps with the ground truth, calculated as the overlap area divided by the combined area. It's one of the most widely used metrics for scoring object detection accuracy.
Inverse Kinematics
A method for calculating the joint angles or movements a robot needs to position its end effector (such as a hand or gripper) at a desired location. Unlike forward kinematics, which computes position from known joint angles, inverse kinematics works backward from a target position to determine how the joints must move to reach it.
K
Keypoint Detection
A computer vision technique that identifies specific, predefined points of interest on an object such as joints on a human body, corners on a vessel, rather than its full outline. It underpins tasks like pose estimation that need precise structural understanding, not just a bounding box.
L
LiDAR
LiDAR (Light Detection and Ranging) uses pulsed laser light to measure distances and build a precise 3D point cloud of the surrounding environment. It's a core sensor for autonomous systems, and one that's notoriously hard to simulate accurately which is exactly why high-fidelity LiDAR simulation matters.
Locomotion
The methods and mechanisms by which a robot moves through its environment, such as walking, rolling, crawling, or flying. In physical AI, locomotion involves coordinating motors, joints, and control policies to achieve stable, efficient movement across varied terrain.
Loiter Time
Loiter time is the duration an aircraft, typically a drone, can remain airborne over a target area before needing to return or refuel. Longer loiter time generally means more persistent ISR coverage, but also more operating conditions a system needs to handle reliably over an extended mission.
Long-tail Scenarios
Long-tail scenarios are the large number of individually rare situations that, collectively, account for a meaningful share of real-world conditions an AI system will encounter.
Curious what your policy actually does in the field?
Tell us what you're building. We'll get you access to Manifold and Stardust.
