Skip to main content

Robotics Glossary

A

Action Space: Set of all possible actions a robot can take. Can be discrete (e.g., move forward, turn left) or continuous (e.g., joint velocities).

Actuator: Device that produces motion (e.g., motors, hydraulic cylinders, pneumatic pistons).

AGV (Automated Guided Vehicle): Mobile robot that follows markers or wires on the floor, or uses vision or lasers for navigation.

AMR (Autonomous Mobile Robot): Robot that navigates using sensors and onboard decision-making, without fixed paths.

Anthropomorphic: Resembling human form or characteristics (e.g., humanoid robots).

B

Baseline: Distance between two stereo camera sensors, affecting depth estimation accuracy.

Behavior Tree: Hierarchical structure for robot decision-making, organizing tasks into sequences, selectors, and parallel nodes.

Bipedal Locomotion: Walking on two legs, requiring advanced balance control (ZMP, COM management).

Bounding Box: Rectangle (2D) or cuboid (3D) enclosing a detected object.

C

Calibration: Process of determining accurate parameters (camera intrinsics, joint offsets, sensor biases).

Center of Mass (COM): Point where an object's mass is concentrated, critical for balance control.

Collision Detection: Identifying when robot links intersect with obstacles or self-collide.

Configuration Space (C-Space): Space of all possible joint configurations of a robot.

D

Degrees of Freedom (DOF): Number of independent ways a robot can move. Humanoid arm typically has 7 DOF.

Depth Image: Image where each pixel represents distance to camera (e.g., from RealSense, stereo cameras).

Denavit-Hartenberg (DH): Standard convention for describing robot kinematics using 4 parameters per joint.

Differential Drive: Two-wheel drive system where speed difference creates turning motion.

Digital Twin: Virtual replica of physical robot/environment, synchronized in real-time.

Domain Randomization: Varying simulation parameters (lighting, textures) to improve sim-to-real transfer.

E

Edge Computing: Processing data on local devices (Jetson Orin) rather than cloud, reducing latency.

Embodied AI: AI systems with physical presence that interact with the real world.

End-Effector: Terminal device of a robot arm (gripper, tool, camera).

Exteroceptive Sensor: Sensor that measures external environment (camera, LiDAR, ultrasonic).

F

Force Closure: Grasp configuration that can resist any external force/torque.

Forward Kinematics (FK): Computing end-effector pose from joint angles.

FPS (Frames Per Second): Rate at which images are processed (e.g., 30 FPS for real-time vision).

Friction Cone: Cone of valid contact forces determined by friction coefficient.

G

Gait: Pattern of leg movement during locomotion (e.g., walking, running, trotting).

Gazebo: Open-source robot simulator supporting ROS integration.

GPU Acceleration: Using graphics processors (NVIDIA RTX, Jetson) for parallel computation (perception, physics).

Grasp Planning: Determining where and how to grasp an object for stable manipulation.

H

Holonomic: Robot that can move in any direction instantly (e.g., omnidirectional wheels). Most robots are non-holonomic.

Humanoid: Robot with human-like body structure (head, torso, two arms, two legs).

I

IMU (Inertial Measurement Unit): Sensor measuring acceleration and angular velocity (gyroscope + accelerometer).

Impedance Control: Control strategy creating virtual mass-spring-damper for compliant motion.

Inverse Kinematics (IK): Computing joint angles to achieve desired end-effector pose.

Isaac ROS: NVIDIA's GPU-accelerated ROS 2 packages for perception and navigation.

Isaac Sim: NVIDIA's photorealistic robot simulator built on Omniverse.

J

Jacobian: Matrix relating joint velocities to end-effector velocities, used in IK and control.

Jetson: NVIDIA's edge AI computing platform (Nano, Xavier, Orin) for robotics.

K

Kinematic Chain: Series of rigid links connected by joints.

Kinematics: Study of motion without considering forces.

L

LiDAR (Light Detection and Ranging): Sensor that measures distances using laser pulses, creating 3D point clouds.

LLM (Large Language Model): AI model trained on text (GPT-4, Llama, Claude) used for robot task planning.

Localization: Determining robot's position in an environment (AMCL, particle filter, VSLAM).

M

Manipulation: Moving and interacting with objects using robot arms/grippers.

MATLAB: Programming environment commonly used for robotics algorithms and simulation.

MDP (Markov Decision Process): Mathematical framework for decision-making where outcomes depend only on current state.

MoveIt: Motion planning framework for ROS/ROS 2, using OMPL planners.

Multimodal: Combining multiple data types (vision + language, RGB + depth).

N

Nav2: ROS 2 navigation stack for mobile robots (path planning, obstacle avoidance).

NLU (Natural Language Understanding): Converting human language to structured robot commands.

O

Observation Space: Set of all possible sensor readings a robot can receive.

Occupancy Grid: 2D map where each cell indicates free/occupied/unknown space.

Odometry: Estimating position by integrating wheel encoder or visual feature measurements.

OMPL (Open Motion Planning Library): Library of sampling-based motion planning algorithms (RRT, PRM).

Omniverse: NVIDIA's platform for 3D simulation and collaboration, foundation for Isaac Sim.

P

PhysX: NVIDIA's physics engine used in Isaac Sim for realistic dynamics.

PID Controller: Proportional-Integral-Derivative controller for tracking reference signals.

Pointcloud: Set of 3D points representing object/environment surface (from LiDAR, depth cameras).

Policy: Function mapping observations to actions in reinforcement learning.

Proprioceptive Sensor: Sensor measuring robot's internal state (joint encoders, force/torque sensors).

Q

Quaternion: 4-element representation of 3D rotation (x, y, z, w), avoiding gimbal lock.

QP (Quadratic Programming): Optimization method for whole-body control with multiple objectives.

R

RGBD: Color (RGB) + Depth image format from cameras like RealSense.

Reinforcement Learning (RL): Training robots through trial-and-error with reward signals.

Robot Operating System (ROS): Middleware providing communication, tools, and libraries for robotics.

RRT (Rapidly-exploring Random Tree): Sampling-based motion planning algorithm.

S

SLAM (Simultaneous Localization and Mapping): Building map while tracking robot position.

SDF (Simulation Description Format): XML format for describing robot and world models in Gazebo.

Semantic Segmentation: Labeling each pixel in an image with object class.

Sensor Fusion: Combining data from multiple sensors for improved estimates.

Sim-to-Real: Transferring policies learned in simulation to physical robots.

Stereo Vision: Using two cameras to estimate depth through triangulation.

T

TCP (Tool Center Point): Reference point on robot end-effector for control.

TensorRT: NVIDIA's inference optimizer for deep learning models (INT8, FP16 quantization).

TF (Transform): ROS coordinate frame transformation system (TF2 in ROS 2).

Trajectory: Time-parameterized path defining position, velocity, acceleration.

URDF (Unified Robot Description Format): XML format for describing robot structure and properties.

V

VLA (Vision-Language-Action): System integrating visual perception, language understanding, and robot control.

Visual Servoing: Using visual feedback to control robot motion (IBVS, PBVS).

VSLAM (Visual SLAM): SLAM using cameras instead of LiDAR (ORB-SLAM, Isaac Visual SLAM).

W

Whole-Body Control: Coordinating all robot joints to achieve multiple objectives simultaneously.

Workspace: Set of all positions reachable by robot end-effector.

Z

Zero Moment Point (ZMP): Point on ground where net moment from gravity and inertia is zero, used for bipedal balance.

Zero-Shot: Model generalizing to new tasks/objects without specific training examples.