Skip to main content

Glossary

Glossary: Physical AI & Humanoid Robotics​

This glossary defines technical terms used throughout the textbook. Terms are organized by category for easy reference.

Embodied Intelligence & Physical AI​

Embodied Intelligence​

The capability of an AI system to understand and act on the physical world through a robotic body, integrating perception, reasoning, and control in real-time.

Physical AI​

A branch of artificial intelligence focused on systems that interact with and control the physical world, combining digital learning with hardware actuation.

Sim-to-Real Transfer​

The process of training a robot control system in simulation and successfully deploying it to a real robot, accounting for differences between simulation and reality (the "reality gap").

Reality Gap​

The discrepancy between simulated robot behavior and actual physical behavior, caused by imperfect physics modeling, sensor inaccuracy, and environmental unpredictability.

ROS 2 (Robot Operating System 2)​

ROS 2​

Robot Operating System 2; a flexible middleware framework for building modular robot applications with support for real-time constraints, security, and distributed computing.

Node​

A ROS 2 process that performs a specific computational task (e.g., reading a sensor, controlling a motor, processing vision data). Nodes communicate via topics and services.

Topic​

A named bus for communication between ROS 2 nodes. Nodes can publish messages to a topic or subscribe to receive messages, enabling asynchronous, decoupled communication.

Message​

A data structure published to a topic or sent via a service. Examples: sensor_msgs/Image (camera data), geometry_msgs/Twist (velocity commands).

Publisher​

A ROS 2 object in a node that publishes (sends) messages to a topic at regular intervals.

Subscriber​

A ROS 2 object in a node that listens to messages published on a topic and triggers a callback function when a message arrives.

Service​

A request-response communication pattern in ROS 2. One node (server) waits for requests, processes them, and sends back a response. Example: spawn_entity service in Gazebo.

Action​

A asynchronous request-response pattern in ROS 2 supporting long-running tasks with feedback. Unlike services, actions can provide intermediate updates and allow cancellation.

Action Server​

A ROS 2 node component that receives action requests, executes them, and returns results with periodic feedback (e.g., a navigation server executing move commands).

Action Client​

A ROS 2 component that requests an action from an action server, receives feedback during execution, and gets the final result.

Package​

A directory containing ROS 2 code, configuration files, and metadata (package.xml). A package is the unit of organization and distribution in ROS 2.

DDS (Data Distribution Service)​

The middleware protocol underlying ROS 2 for reliable, efficient, real-time communication across a network without a centralized message broker.

rclpy​

The Python client library for ROS 2, allowing Python developers to create nodes and communicate with other ROS 2 components.

rclcpp​

The C++ client library for ROS 2, used for performance-critical real-time applications.

Launch File​

An XML file that specifies which ROS 2 nodes to start, their parameters, and how they connect. Simplifies starting complex multi-node systems with one command.

Parameter​

A configuration value in ROS 2 that can be set at runtime. Parameters allow nodes to adapt behavior without code changes.

Parameter Server​

A ROS 2 service that stores and distributes parameters across nodes.

rqt​

A Qt-based GUI framework for ROS 2 introspection and control. Tools like rqt_graph, rqt_topic, and rqt_image_view help visualize and debug running systems.

rostopic​

A ROS 2 command-line tool for inspecting and interacting with topics (e.g., rostopic echo, rostopic list).

rosservice​

A ROS 2 command-line tool for calling services (e.g., rosservice call).

rosnode​

A ROS 2 command-line tool for inspecting nodes (e.g., rosnode list, rosnode info).

Robotics Fundamentals​

Humanoid Robot​

A robot designed with a body structure resembling a human, typically with two arms, two legs, a torso, and a head, enabling it to interact with human-designed environments.

Kinematics​

The branch of mechanics describing the motion of a robot without considering forces. Forward kinematics calculates end-effector position from joint angles; inverse kinematics solves the reverse.

Dynamics​

The branch of mechanics describing how forces and torques cause motion in a robot. Essential for control systems that must balance forces and accelerations.

Joint​

A movable connection between two links of a robot. Common types: revolute (rotation), prismatic (linear), continuous (unlimited rotation).

A rigid body segment of a robot connected to other links via joints. Example: upper arm, forearm, gripper.

End-Effector​

The tool or appendage at the end of a robotic arm (e.g., gripper, camera, laser pointer).

Actuator​

A motor or solenoid that converts electrical or pneumatic energy into mechanical motion to move robot joints.

Sensor​

A device that measures physical quantities (e.g., position, velocity, force, light, sound) and converts them to electrical signals for processing.

URDF (Unified Robot Description Format)​

An XML format for describing robot structure, including links, joints, sensors, and visual/collision properties. Used by ROS 2 and simulation engines.

SDF (Simulation Description Format)​

An XML format for describing robotic worlds, including robots, objects, physics, and lighting. More powerful than URDF; used in Gazebo.

SLAM (Simultaneous Localization and Mapping)​

A technique for a robot to build a map of its environment while simultaneously determining its own position within that map, essential for autonomous navigation.

Localization​

The process of determining a robot's position and orientation in a known environment (or map).

Mapping​

The process of creating a map of an environment (typically using LiDAR or camera data) for navigation and obstacle avoidance.

Visual Odometry (VO)​

Estimating a robot's motion and position using only camera images, without requiring external markers or prior maps.

Visual SLAM (VSLAM)​

A variant of SLAM using camera images as the primary sensor, combining visual odometry with loop closure and mapping.

Path Planning​

The algorithm for computing a safe, efficient path from a robot's current position to a goal while avoiding obstacles.

Inverse Kinematics (IK)​

Calculating the joint angles required to reach a desired end-effector position and orientation.

Forward Kinematics (FK)​

Computing the end-effector position and orientation from given joint angles.

Collision Avoidance​

Techniques to prevent a robot from colliding with obstacles or itself during motion.

Simulation & Physics​

Gazebo​

A 3D simulation environment for robotics with built-in physics engines, sensor simulation, and tight integration with ROS 2. Used to test robot behavior before physical deployment.

Physics Engine​

Software that simulates real-world physics (gravity, collisions, friction) in a virtual environment. Gazebo supports ODE, Bullet, DART, and Simbody.

ODE (Open Dynamics Engine)​

An open-source physics engine used in Gazebo for simulating rigid body dynamics and collisions.

Bullet Physics​

An open-source physics engine known for accuracy and speed, usable with Gazebo.

DART Physics​

A physics engine emphasizing accurate dynamics simulation, used in Gazebo and research.

Mesh​

A 3D geometric representation of an object made of vertices, edges, and faces. Used for collision detection and visualization.

Inertia​

A property describing how mass is distributed in an object, affecting rotational dynamics. Every link in a URDF must specify inertia.

Gravity​

Downward acceleration (9.81 m/s² on Earth) simulated in physics engines to create realistic falling and support reactions.

Friction​

Resistance to motion between two surfaces. Gazebo allows tuning friction coefficients.

Collision​

Detection and response when two objects physically intersect, critical for realistic simulation.

RGB-D Camera​

A camera providing both color (RGB) and depth (D) information, essential for 3D vision tasks.

Point Cloud​

A set of 3D points representing a sensed environment, typically from LiDAR or RGB-D cameras.

LiDAR (Light Detection and Ranging)​

A sensor using laser pulses to measure distances to objects, producing a point cloud. Essential for SLAM and obstacle detection.

IMU (Inertial Measurement Unit)​

A sensor measuring linear acceleration and angular velocity, used for tracking robot orientation and motion.

Depth Sensor​

A camera (like Intel RealSense) providing depth information at every pixel, useful for 3D vision and obstacle avoidance.

NVIDIA Isaac Platform​

NVIDIA Isaac Sim​

A photorealistic simulation environment for robots, built on NVIDIA Omniverse, supporting synthetic data generation, hardware-accelerated physics, and AI-based perception training.

NVIDIA Isaac ROS​

A set of ROS 2 packages providing hardware-accelerated perception and control, optimized for NVIDIA Jetson processors and data center GPUs.

Omniverse​

NVIDIA's platform for 3D simulation and collaboration, providing rendering, physics, and AI capabilities for robotics simulation.

Nucleus​

NVIDIA's asset library and content management system, providing 3D models, materials, and environments for use in Isaac Sim.

Synthetic Data​

Artificially generated data (images, sensor readings) from simulation, used to train machine learning models without requiring real hardware.

Hardware-Accelerated Perception​

Using GPU acceleration for real-time image processing, object detection, and SLAM, critical for edge deployment.

Visual Odometry (Isaac ROS)​

A hardware-accelerated visual odometry implementation in Isaac ROS for real-time camera-based localization.

VPX (Voxel Perception Engine)​

An NVIDIA Isaac ROS module for GPU-accelerated 3D perception and scene understanding.

Vision & Computer Vision​

Object Detection​

Identifying and localizing objects in an image, returning bounding boxes and class labels. Example: YOLO, Faster R-CNN.

YOLO (You Only Look Once)​

A real-time object detection algorithm, commonly used in robotics for fast detection of objects in camera feeds.

Semantic Segmentation​

Classifying every pixel in an image into semantic categories (person, car, tree, etc.), useful for scene understanding.

Instance Segmentation​

Distinguishing individual objects of the same class (e.g., detecting and separating two cups in an image).

Optical Flow​

Estimating pixel-level motion between consecutive frames, used for visual odometry and motion detection.

Feature Detection​

Identifying distinctive points or regions in an image (keypoints) used for matching images and visual odometry.

Image Processing​

Algorithms for enhancing, filtering, or transforming images (e.g., edge detection, histogram equalization).

Vision-Language-Action (VLA) Systems​

Vision-Language-Action (VLA)​

A multimodal AI system that combines vision (image input), language (natural language commands), and action (robot motor commands) to enable robots to understand and execute human instructions.

Multimodal AI​

AI systems that process multiple types of input (vision, language, audio, proprioception) to make decisions or generate outputs.

Large Language Model (LLM)​

A deep learning model trained on vast amounts of text, capable of understanding and generating human language. Examples: GPT-4, Claude, Llama.

Prompt Engineering​

Crafting input prompts to an LLM to elicit desired behavior, critical for robotics where the model must generate executable actions.

Grounding​

Connecting abstract language concepts to concrete perceptual or motor representations, e.g., linking the word "grasp" to specific robot joint commands.

Action Space​

The set of all possible actions a robot can execute (e.g., joint velocities, gripper commands).

Semantic Understanding​

Extracting meaning from language or images beyond surface-level patterns, essential for generalization.

Transfer Learning​

Adapting a model trained on one task to perform a different task, reducing data and computation required.

Few-Shot Learning​

Learning to perform a new task from only a few examples, important for robots to adapt to new scenarios.

Speech & Audio​

Automatic Speech Recognition (ASR)​

Converting spoken language (audio) into text, essential for voice-controlled robots.

Whisper (OpenAI)​

An open-source ASR model capable of transcribing speech in multiple languages, commonly used in VLA robotics.

Text-to-Speech (TTS)​

Converting text into spoken audio, used for robots to provide voice feedback.

ReSpeaker​

A microphone array (e.g., 6-mic circular array) optimized for voice input, commonly used in robotics for directional audio.

Voice Commands​

Human instructions given verbally to a robot, processed by ASR and then interpreted by an LLM.

Audio Processing​

Techniques for filtering, enhancing, or analyzing audio signals (e.g., noise reduction, beamforming).

Hardware & Platforms​

Jetson Orin Nano​

NVIDIA's edge AI processor for robotics, providing GPU acceleration for perception and inference on a small, power-efficient platform.

Jetson Orin Nano Developer Kit​

A single-board computer with NVIDIA Orin Nano GPU, used for developing robotics applications with hardware acceleration.

Unitree G1​

A humanoid robot platform developed by Unitree, featuring full-body control, balance, and multimodal sensing for research and development.

Boston Dynamics Atlas​

A state-of-the-art humanoid research platform, showcasing advanced mobility and manipulation.

Tesla Optimus (Bot)​

A humanoid robot under development by Tesla, targeting general-purpose manipulation tasks.

RealSense D435i​

An Intel RGB-D depth camera commonly used in robotics for 3D vision and obstacle detection.

Ubuntu 22.04 LTS​

The recommended operating system for ROS 2 Humble, providing a stable, long-term support base for robotics development.

Docker​

A containerization platform for packaging robots software and dependencies in isolated, reproducible environments.

Machine Learning & AI​

Deep Learning​

Machine learning using neural networks with multiple layers, powerful for vision and language tasks.

Convolutional Neural Network (CNN)​

A deep learning architecture specialized for processing grid-like data (images), using convolutional layers to extract spatial features.

Transformer​

A neural network architecture based on attention mechanisms, effective for sequence processing (language, time-series).

Attention Mechanism​

A technique for models to focus on relevant parts of input, foundational for modern LLMs and vision transformers.

Fine-Tuning​

Adapting a pre-trained model to a specific task with task-specific data, more efficient than training from scratch.

Model Quantization​

Reducing the precision of model weights (e.g., from 32-bit floats to 8-bit integers) to decrease memory and computation, enabling edge deployment.

Inference​

Running a trained model on new data to make predictions or generate outputs, as opposed to training.

Latency​

The time delay between input and output, critical for real-time robotic control (must be under 100ms typically).

Throughput​

The number of samples processed per unit time, important for high-frequency sensor data processing.

Control & Systems​

Control System​

A system that regulates a robot's behavior by receiving sensor feedback and adjusting actuator commands.

Feedback Control​

Using sensor measurements (feedback) to adjust motor commands, enabling stability and accuracy.

Feedforward Control​

Pre-planning motor commands based on a model without real-time feedback; less robust but potentially faster.

Closed-Loop Control​

A control system with feedback, allowing real-time adjustments; more stable and adaptive.

Open-Loop Control​

A control system without feedback, simply executing pre-planned commands; sensitive to disturbances.

PID Control​

Proportional-Integral-Derivative control, a classic feedback control algorithm adjusting output based on error, error integral, and error rate.

Stability​

The property of a control system where disturbances don't cause unbounded behavior; critical for safe robot operation.

Tracking​

The ability of a robot to follow a desired trajectory or target position with high accuracy.

Response Time​

The time for a robot to react to a command or stimulus, important for real-time interaction.

Bandwidth​

The frequency range of disturbances or commands a control system can effectively handle.

Safety & Ethics​

Safety-Critical System​

A system where failures could cause injury or property damage, requiring rigorous validation and redundant safeguards.

Collision Avoidance​

Techniques preventing a robot from contacting objects or people during operation.

Force Limiting​

Hardware or software limiting the maximum force a robot can apply, protecting people and objects.

Emergency Stop (E-Stop)​

A button or mechanism to immediately halt robot motion in case of danger.

Validation​

Ensuring a system meets intended requirements and behaves safely in real-world conditions.

Risk Mitigation​

Strategies to reduce the likelihood and severity of potential hazards (e.g., redundancy, failsafes, training).

Development & Software Engineering​

Package Manager​

Software (like apt, pip, conda) that installs and manages libraries and dependencies.

Colcon​

A build tool for ROS 2 packages, replacing catkin, supporting C++ and Python packages in a single workspace.

Build System​

Software (like CMake) managing compilation, linking, and packaging of source code.

Continuous Integration (CI)​

Automatically building and testing code on every commit, catching errors early.

Version Control​

Tracking changes to code over time, typically using Git and GitHub.

API (Application Programming Interface)​

The interface a library or service provides, specifying how to call functions and pass data.

JSON (JavaScript Object Notation)​

A human-readable data format commonly used for configuration and data exchange.

YAML (YAML Ain't Markup Language)​

A human-readable data format used for ROS 2 configuration files (launch files, parameters).

XML (eXtensible Markup Language)​

A structured data format used for URDF, SDF, and ROS 2 launch files.

Markdown​

A lightweight text format used for documentation, including this textbook.

Git​

A version control system for tracking code changes and collaborating with others.

GitHub​

A cloud platform for hosting Git repositories, enabling collaboration and code sharing.

Acronyms & Abbreviations​

AcronymFull Form
ASRAutomatic Speech Recognition
CNNConvolutional Neural Network
DDSData Distribution Service
FKForward Kinematics
GPUGraphics Processing Unit
IKInverse Kinematics
IMUInertial Measurement Unit
JSONJavaScript Object Notation
LiDARLight Detection and Ranging
LLMLarge Language Model
NLPNatural Language Processing
ODEOpen Dynamics Engine
PIDProportional-Integral-Derivative
ROSRobot Operating System
RGB-DRed-Green-Blue-Depth
SLAMSimultaneous Localization and Mapping
SDFSimulation Description Format
TTSText-to-Speech
URDFUnified Robot Description Format
VOVisual Odometry
VSLAMVisual SLAM
VLAVision-Language-Action
YAMLYAML Ain't Markup Language

Further Learning​

For deeper understanding of robotics and AI concepts, consult:

Last Updated: 2025-12-10 Maintained By: Textbook Team

Textbook Assistant

Ask me anything about the textbook...