AutomotiveCloudWatch

Glossary · Robotics and Physical AI

Perception, Navigation and Planning

129 terms in the Robotics and Physical AI glossary, each with its one-line definition.

129 terms, A to Z

  • 3D Scene Graph Layered map that stores a space as buildings, rooms, places and objects plus their relations, so planners and language models can query it.
  • 6D Pose Estimation Working out an object's 3D position and 3D orientation from camera images so a robot can grasp or insert it correctly.
  • A Star Search A* Classic shortest-route algorithm that explores the most promising map cells first by adding distance traveled to an estimate of distance remaining.
  • Active Perception Moving a robot's sensors on purpose, such as tilting a wrist camera, to get a better view before acting.
  • AnyGrasp Grasp detection system that proposes many gripper grasps from a depth view of cluttered, unseen objects and tracks them as things move.
  • Artificial Potential Field APF Navigation method where the goal attracts the robot and obstacles repel it, simple and fast but prone to getting stuck in local minima.
  • Bayes Filter General recipe for tracking an uncertain state by alternately predicting after motion and correcting with each sensor reading.
  • Behavior Tree BT Way of structuring robot decisions as a tree of small reusable actions and checks that report success or failure upward.
  • Bird's-Eye View Representation BEV Top-down map of a robot's or vehicle's surroundings built from cameras and other sensors, showing obstacles, lanes and free space.
  • Bundle Adjustment Joint optimization of all camera poses and 3D points so a reconstruction lines up best across many images.
  • Camera Calibration Measuring a camera's focal length, lens distortion and mounting position so pixels can be converted into real-world positions.
  • Cartographer Open-source real-time 2D and 3D lidar SLAM system from Google, long used in ROS to map buildings and warehouses.
  • Collision Checking Test a planner runs to see whether a robot pose or path would hit obstacles or its own body.
  • Computer Vision CV Field of getting computers to understand images and video, now the main way most object-handling robots sense their world.
  • Configuration Space C-space Space where each robot pose is one point with one axis per joint, turning arm motion planning into moving a point around obstacles.
  • Conflict-Based Search CBS Multi-robot route planning algorithm that plans each robot alone and then resolves pairwise conflicts as they appear.
  • Costmap Grid map where each cell holds a cost for driving through it, used by navigation planners to find short, safe paths.
  • Covariant Hamiltonian Optimization for Motion Planning CHOMP Trajectory optimizer that bends a rough, possibly colliding arm path away from obstacles while keeping it smooth.
  • Coverage Path Planning Planning a route that passes over every part of an area, as needed by cleaning, mowing, painting and inspection robots.
  • cuRobo NVIDIA GPU-accelerated library that plans collision-free robot arm motions in tens of milliseconds.
  • cuVSLAM NVIDIA's GPU-accelerated visual SLAM library that tracks robot position from stereo cameras and an optional IMU.
  • D Star Lite D* Lite Path planner that quickly repairs its route when the robot discovers the map was wrong, instead of replanning from scratch.
  • Data Association Deciding which sensor measurement belongs to which landmark, object or track, a common failure point in mapping and tracking.
  • Dead Reckoning Estimating position by adding up known movements from a starting point, which drifts until corrected by an outside reference.
  • Depth Anything Family of AI models that estimate relative depth from a single ordinary photo, giving robots depth hints without a depth sensor.
  • Depth Estimation Working out how far away each point in an image is, from stereo cameras, active sensors or AI models.
  • Dex-Net Dexterity Network UC Berkeley project that trained grasp-quality networks on millions of synthetic grasps to pick unfamiliar items from depth images.
  • Dijkstra's Algorithm Shortest-path algorithm that always expands the closest unexplored node, guaranteeing the best route but searching in all directions.
  • Dubins Path Shortest route between two poses for a forward-only vehicle with a minimum turning radius, made of arcs and straight lines.
  • DUSt3R Neural network that reconstructs 3D geometry and camera poses from a few ordinary photos in one pass.
  • Dynamic Window Approach DWA Local planner that samples reachable speeds and turns, predicts each briefly ahead and picks the best obstacle-free option.
  • Elevation Mapping Building a height map of the terrain around a robot, which legged robots use to choose safe footholds.
  • Embodied Question Answering EQA Task where an agent explores a building to answer a question, combining navigation, vision and language.
  • Extended Kalman Filter EKF Kalman filter for nonlinear systems that linearizes around the current estimate, the workhorse for fusing wheel, IMU and GPS data.
  • Factor Graph Graph linking unknown states to the measurements that constrain them, solved to find values that best fit all the evidence.
  • FAST-LIO Fast open-source lidar-inertial odometry from the University of Hong Kong, widely used on drones, legged robots and mapping rigs.
  • Feature Matching Finding the same distinctive spots in two images so software can recover camera motion or 3D structure.
  • Finite State Machine FSM Behavior described as a short list of states and the events that switch between them, common in industrial robot cells.
  • FoundationPose NVIDIA AI model that estimates and tracks the 6D pose of new objects from a CAD model or a few photos, without retraining.
  • FoundationStereo NVIDIA AI model that computes accurate depth from a stereo camera pair, even on shiny or transparent surfaces.
  • Frontier-Based Exploration Mapping strategy where a robot repeatedly drives to the boundary between known and unknown space until nothing unknown remains.
  • Global Planner Navigation component that plans the whole route from the robot's position to the goal using the full map.
  • Grasp Detection AI that proposes where a gripper should grab directly from camera or depth data, without first identifying the object.
  • Grasp Planning Deciding where and how a gripper holds an object, including contact points, approach direction and force.
  • Grounding DINO Open-vocabulary detector that draws boxes around objects described in typed text, often paired with a segmentation model.
  • GTSAM Open-source C++ and Python library from Georgia Tech for solving SLAM and sensor fusion as factor graphs.
  • Human Pose Estimation Software that finds a person's body joints in images to keep a safe distance, read gestures or learn from human motion.
  • Hybrid A Star Hybrid A* Route planner for car-like robots that searches over position and heading so paths are actually drivable, including reversing.
  • Image Segmentation Labeling every pixel in an image by what it belongs to, giving exact outlines for grasping or walkable floor.
  • Iterative Closest Point ICP Aligning two 3D scans by repeatedly pairing nearest points and shifting one scan until they fit.
  • Kalman Filter KF Method that blends a prediction with noisy sensor readings to estimate a state such as robot position, repeating many times a second.
  • Keepout Zone Area drawn on a robot's map that it must never enter, even when sensors see open floor there.
  • Keypoint Detection Finding a few meaningful points in an image, such as a mug handle or towel corner, to summarize where to act.
  • Kidnapped Robot Problem Test of whether a robot can relocate itself after being moved without warning, making its prior estimate wrong.
  • Kinodynamic Planning Motion planning that respects speed, acceleration and force limits so the result can actually be followed by a real robot.
  • Lidar Odometry and Mapping LOAM Influential method that estimates lidar motion and builds a 3D map by matching edge and plane points between scans.
  • LiDAR SLAM Building a map and tracking position at once with a laser scanner, robust in the dark and the default for warehouse robots.
  • Local Planner Navigation component that steers moment to moment, following the global route while dodging unmapped obstacles.
  • Localization Robot working out where it is and which way it faces on a map it already has, its you-are-here dot.
  • Loop Closure Recognizing a return to a previously visited place so a mapping robot can correct accumulated drift across its whole map.
  • Mapping Robot building a usable representation of its surroundings, such as a grid, point cloud or labeled map.
  • Model Predictive Path Integral Control MPPI Sampling controller that simulates thousands of short futures, scores them and blends the best, many times a second.
  • Monte Carlo Localization MCL Particle-based localization that keeps many position guesses and keeps those matching the sensors until they converge.
  • Motion Planning Computing how all of a robot's joints should move from start to goal without collisions or exceeding limits.
  • Multi-Agent Path Finding MAPF Planning routes for many robots together so they share a floor without collisions or deadlocks.
  • Navigation Foundation Model Single AI model trained on many robots' driving and walking data that can steer a new robot toward goals without retraining.
  • Navigation Stack Full software bundle that lets a mobile robot move on its own, from localization and maps to planning and recovery.
  • Nonholonomic Constraint Limit on instantaneous motion, like a car that cannot slide sideways, that still lets the robot reach any pose by maneuvering.
  • Normal Distributions Transform NDT Scan matching method that models lidar points per grid cell as statistics and finds the best-fitting pose, popular outdoors.
  • nvblox NVIDIA library that builds a real-time 3D map with obstacle distances on the GPU from depth cameras or lidar.
  • Object Detection Finding objects in an image and labeling each with a box, class and confidence score.
  • Object Tracking Following objects across video frames while keeping their identities, such as a part on a conveyor or a walking person.
  • Object-Goal Navigation ObjectNav Task where a robot must find a named type of object, such as a chair, in an unseen building by exploring.
  • Obstacle Avoidance Robot noticing unexpected or moving things in its way and steering around them in real time.
  • Occupancy Grid Floor map of small cells each marked free, occupied or unknown, the standard map for wheeled indoor robots.
  • OctoMap Open-source library for 3D maps of occupied, free and unknown space stored in a tree of nested cubes.
  • Odometry Robot estimating its own motion from wheel turns or shifting camera images, smooth but drifting over time.
  • Open Motion Planning Library OMPL Free library of sampling-based motion planners and the default planning engine inside MoveIt.
  • Open-Vocabulary Perception Finding or outlining objects described in everyday words rather than only a fixed list learned in training.
  • OpenCV Free, widely used library of image and video tools for filtering, calibration, feature finding and marker detection.
  • Optical Flow Apparent motion of each part of an image between video frames, used to sense movement and extract actions from videos.
  • Optimal Reciprocal Collision Avoidance ORCA Method where moving robots or agents each choose a velocity that shares the avoidance effort equally with others.
  • ORB-SLAM Widely used open-source visual SLAM system using ORB image features with mono, stereo or depth cameras.
  • OWL-ViT Google open-vocabulary detector that finds objects from a typed description or example image.
  • Particle Filter Estimator that tracks many weighted guesses and resamples them as evidence arrives, able to hold several hypotheses at once.
  • Path Planning Finding an obstacle-free route from a start to a goal, usually for a robot moving across a floor or through the air.
  • Perspective-n-Point PnP Computing a camera's position and orientation from known 3D points and where they appear in the image.
  • Planning Domain Definition Language PDDL Standard language for writing planning problems as objects, actions, preconditions, effects and goals for automated planners.
  • Point Cloud Collection of 3D dots measured on surfaces by lidar or depth cameras, used for mapping, collision checks and grasping.
  • Point Cloud Library PCL Free toolkit for filtering, segmenting and aligning 3D point clouds, long part of ROS perception.
  • Point Cloud Registration Finding the shift and rotation that best overlaps two 3D scans, used to build maps, track motion or match parts.
  • Point Tracking Following any chosen spot through a video, even when briefly hidden, often used to describe object motion for robot learning.
  • Point-Goal Navigation PointNav Benchmark task where a robot must reach coordinates given relative to its start using only its own sensors.
  • Pose Estimation Working out where something is and which way it faces, whether an object, a person or the camera.
  • Pose Graph Optimization Correcting a map by adjusting all recorded robot poses until the motion links between them agree best.
  • Probabilistic Roadmap PRM Planner that scatters random collision-free poses, links nearby ones into a network and searches it for each query.
  • Pure Pursuit Steering method that keeps aiming at a point a set distance ahead on the path, giving smooth, easy-to-tune driving.
  • Random Sample Consensus RANSAC Fitting a model to noisy data by repeatedly fitting random subsets and keeping the fit most points agree with.
  • Rapidly-Exploring Random Tree RRT Sampling planner that grows a tree toward random points, quickly finding paths even for robots with many joints.
  • Recovery Behavior Preset action such as clearing the map, spinning, backing up or waiting that a robot tries when stuck.
  • Robot Perception Software that turns raw camera, lidar, depth and touch readings into an understanding of the robot's surroundings.
  • RRT Star RRT* Asymptotically optimal RRT variant that keeps rewiring its tree so paths get shorter the longer it runs.
  • RTAB-Map Free ROS-friendly SLAM package that builds 3D maps and tracks position with depth cameras, stereo or lidar.
  • Sampling-Based Planning Planning by testing random poses, keeping collision-free ones and linking them, the standard approach for robot arms.
  • Segment Anything Model SAM Meta AI model that outlines any object from a click or rough box, even objects it has never seen.
  • Self-Collision One part of a robot hitting another part of itself, such as an arm striking its own torso.
  • Semantic Mapping Adding labels such as kitchen, door or shelf to a robot's map so it can follow meaning-based instructions.
  • Semantic Navigation Getting a robot to a goal described by meaning, such as find a chair, rather than by map coordinates.
  • Signed Distance Field SDF 3D map storing the distance to the nearest surface at every point, negative inside objects, for fast collision checks.
  • Simultaneous Localization and Mapping SLAM Building a map of an unknown place while tracking position within it, the foundation of mobile robot navigation.
  • SLAM Toolbox Free ROS 2 package for 2D lidar mapping and localization, the default mapping tool in Nav2.
  • Social Navigation Moving among people in a natural, polite way, such as keeping comfortable distance and not cutting through groups.
  • State Lattice Planner Planner that searches precomputed, physically feasible motion primitives on a grid of positions and headings.
  • Stereo Matching Computing depth from two side-by-side cameras by finding each point's shift, or disparity, between the images.
  • Structure from Motion SfM Reconstructing a scene's 3D shape and camera positions from overlapping photos, often feeding NeRF and splatting.
  • Task and Motion Planning TAMP Planning both which steps to take and exactly how to move for each, so a multi-step plan is physically possible.
  • Task Planning Ordering high-level steps toward a goal, such as open the door then pick up the cup, without exact joint motions.
  • Topological Map Map of places such as rooms and doorways joined by links, like a subway map, scaling well to large buildings.
  • Trajectory Optimization Computing a motion by minimizing time, energy or jerk subject to limits and obstacles, often polishing a rough path.
  • Traversability Estimation Predicting which ground a robot can safely drive or walk over and how hard it will be, essential outdoors.
  • Unscented Kalman Filter UKF Kalman filter variant that passes chosen sample points through nonlinear equations, often beating the EKF.
  • Vision-Language Navigation VLN Task where a robot follows spoken or written route directions using only what its camera sees.
  • Visual Odometry VO Estimating camera and robot motion by comparing consecutive images, smooth but drifting over time.
  • Visual Place Recognition VPR Recognizing a previously seen place from a camera image despite changes in light, weather, season or angle.
  • Visual Servoing Steering a robot directly from camera images, adjusting until the view looks right, for precise alignment.
  • Visual SLAM vSLAM Building a map and tracking position at once with cameras, popular for drones and low-cost robots.
  • Visual-Inertial Odometry VIO Tracking motion by fusing a camera with an IMU, so each covers the other's weaknesses.
  • Voxel Map 3D grid of small cubes each marked occupied, free or unknown, the 3D version of a floor grid map.
  • You Only Look Once YOLO Family of fast single-pass object detectors that run at video speed on small robot computers.

Full entry for members

Every term has a full entry for members: the plain-words definition, the practitioner note and the typed links between terms.

See membership

ACW Robotics and Physical AI Glossary, edition 2026.11.1. The headword and one line are open to everyone.