Glossary
Comprehensive glossary of terms used throughout the Physical AI & Humanoid Robotics textbook, organized alphabetically.
A
Action Model: Component of a VLA system that translates high-level plans into low-level robot control commands (joint positions, gripper commands, wheel velocities).
Action Space: In reinforcement learning, the set of all possible actions an agent can take (e.g., 12-dimensional continuous space for humanoid joint positions).
Actuator: A mechanical device that converts energy (electrical, hydraulic, pneumatic) into motion, such as motors that drive robot joints.
A2C (Advantage Actor-Critic): A reinforcement learning algorithm that uses both a policy network (actor) and a value network (critic) to improve learning efficiency.
Articulation: A robotic system with joints connecting rigid bodies (links), represented in Isaac Sim as a connected hierarchy with constraints.
Autonomous Navigation: The capability of a robot to plan and execute paths from one location to another without human intervention, avoiding obstacles.
B
Baseline (ROS 2): In robotics, often refers to the distance between stereo camera lenses, affecting depth accuracy; also can mean a reference implementation for comparison.
Bounding Box: A rectangular region in an image that encloses a detected object, typically defined by (x, y, width, height).
C
Calibration (Camera): The process of determining intrinsic parameters (focal length, principal point, distortion) and extrinsic parameters (position, orientation) of a camera for accurate 3D reconstruction.
CLIP (Contrastive Language-Image Pre-training): OpenAI's vision-language model that learns joint embeddings of images and text, enabling zero-shot object detection by matching text queries to image regions.
Closed-Loop Control: A control system that uses feedback from sensors to adjust actions in real-time, correcting for errors and disturbances (contrasts with open-loop control).
Colcon: The build system for ROS 2 packages, replacing catkin from ROS 1, used to compile C++/Python code and manage dependencies.
Contact Sensor: A sensor that detects physical contact between robot parts and the environment, measuring forces or binary contact state (touching/not touching).
Control Loop: The cycle of sensing, computing, and actuating that runs continuously in a robot system, typically at high frequency (100-1000 Hz).
CUDA: NVIDIA's parallel computing platform that enables GPU acceleration for general-purpose computing, essential for deep learning and physics simulation.
D
DDS (Data Distribution Service): The middleware standard used by ROS 2 for real-time publish-subscribe communication between distributed nodes.
Depth Anything: State-of-the-art monocular depth estimation model using vision transformers, trained on millions of RGB-depth pairs to predict dense depth maps from single images.
Depth Map: An image where each pixel value represents the distance from the camera to the corresponding point in the scene, typically in meters.
Digital Twin: A virtual replica of a physical robot or environment used for simulation, testing, and training before deploying to real hardware.
Domain Randomization: A sim-to-real transfer technique that randomizes simulation parameters (physics, visuals, sensor noise) during training to prevent overfitting and improve real-world robustness.
DoF (Degrees of Freedom): The number of independent parameters defining a robot's configuration; a humanoid arm typically has 7 DoF (3 shoulder, 1 elbow, 3 wrist).
E
Embodied AI: Artificial intelligence systems that interact with the physical world through sensors and actuators (robots), requiring grounding of abstract knowledge in physical experience. Also called Physical AI.
End-Effector: The device at the end of a robotic arm that interacts with the environment, such as a gripper, welding tool, or camera.
Episode: In reinforcement learning, a sequence of states, actions, and rewards from an initial state to a terminal state (e.g., robot attempts task from start to success/failure).
F
Focal Length (Camera): The distance from the camera lens to the image sensor, determining the field of view and magnification; key parameter for pinhole camera model.
Force Sensor: A sensor that measures applied forces (and sometimes torques), used in robotic grippers to detect object contact and prevent damage.
Forward Kinematics: Computing the end-effector position and orientation from given joint angles using robot geometry and link transformations.
G
GAE (Generalized Advantage Estimation): An RL algorithm component (parameterized by λ) that balances bias vs variance in advantage estimation for policy gradient updates, used in PPO.
Gazebo: An open-source 3D robot simulator with physics engines (ODE, Bullet, DART), widely used in ROS ecosystem for testing navigation and manipulation.
GPU-Accelerated Physics: Using graphics processing units (GPUs) to parallelize physics simulation (collision detection, constraint solving), achieving 100-1000x speedup over CPU-based simulators.
Grasp Planning: The process of determining where and how a robot gripper should approach and grip an object to achieve a stable grasp.
Grounding Problem: Mapping natural language references (e.g., "the red mug") to specific physical entities in the environment, resolving ambiguities through spatial reasoning and context.
H
Headless Mode: Running a simulator (Isaac Sim, Gazebo) without a graphical user interface (GUI), useful for faster execution on servers or edge devices.
URDF (Unified Robot Description Format): An XML-based file format for describing robot kinematics, dynamics, collision geometry, and visual appearance in ROS.
I
IMU (Inertial Measurement Unit): A sensor that measures linear acceleration and angular velocity, used for estimating robot orientation and motion.
Inverse Kinematics: Computing joint angles required to achieve a desired end-effector position and orientation, typically solved via optimization or analytical methods.
Isaac Lab: NVIDIA's reinforcement learning framework for robotics (formerly Isaac Gym), providing vectorized GPU environments, pre-built tasks, and RL algorithm integration.
Isaac Sim: NVIDIA's robotics simulator built on Omniverse, featuring GPU-accelerated PhysX physics, ray-traced rendering (RTX), USD scene format, and ROS 2 integration.
J
Joint: A connection between two robot links that allows relative motion, categorized as revolute (rotational) or prismatic (linear).
Joint State: The current position, velocity, and effort (torque/force) of all robot joints, published on /joint_states topic in ROS 2.
K
Kinematics: The study of motion without considering forces, including forward kinematics (joint angles → end-effector pose) and inverse kinematics (end-effector pose → joint angles).
L
Latency: The time delay between an event (e.g., sensor reading) and the corresponding response (e.g., actuator command), critical for real-time robot control.
Launch File: A configuration file (Python or XML) in ROS 2 that starts multiple nodes with specified parameters and remappings, simplifying system startup.
LiDAR (Light Detection and Ranging): A sensor that measures distances by emitting laser pulses and measuring time-of-flight, used for 2D/3D mapping and obstacle detection.
Link (URDF): A rigid body component of a robot (e.g., forearm, thigh) defined in URDF with mass, inertia, collision geometry, and visual mesh.
LLM (Large Language Model): Transformer-based neural networks trained on massive text corpora (GPT-4o, Claude, LLaMA) capable of natural language understanding, reasoning, and code generation.
Localization: The process of determining a robot's position and orientation within a known map, often using SLAM (Simultaneous Localization and Mapping).
M
Mermaid: A JavaScript-based diagramming tool that renders diagrams from text descriptions, natively supported in Docusaurus for flowcharts, sequence diagrams, etc.
MoveIt2: The motion planning framework for ROS 2, providing inverse kinematics, collision checking, trajectory generation, and execution for robotic arms.
Multimodal Model: A neural network that processes multiple input types (text, images, audio), such as vision-language models like CLIP or GPT-4o with vision.
N
Nav2 (Navigation 2): The ROS 2 navigation stack providing autonomous navigation capabilities including path planning, obstacle avoidance, and localization.
Node (ROS 2): An independent process in ROS 2 that performs a specific function (e.g., camera driver, motion planner), communicating via topics, services, or actions.
O
Observation Space: In reinforcement learning, the set of all possible observations an agent receives from the environment (e.g., 48-dimensional vector of joint states + IMU data).
Odometry: Estimation of robot position over time based on sensor data (wheel encoders, IMU, visual features), subject to drift without correction.
Omniverse: NVIDIA's platform for 3D collaboration and simulation, providing the foundation for Isaac Sim (rendering, USD support, multi-user workflows).
Open-Loop Control: A control system that executes commands without feedback, assuming perfect execution (contrasts with closed-loop control).
Open-Vocabulary Detection: Object detection that works for ANY text description without retraining, using vision-language models like CLIP (contrasts with closed-vocabulary detectors like YOLO).
P
PhysX 5: NVIDIA's GPU-accelerated rigid body physics engine, capable of simulating thousands of robots in parallel with accurate contact dynamics and articulation solvers.
Physical AI: See Embodied AI. Artificial intelligence systems that perceive and interact with the physical world through sensors and actuators.
Pinhole Camera Model: A mathematical model relating 3D world points to 2D image pixels using focal length and principal point, fundamental to computer vision.
Point Cloud: A set of 3D points representing object surfaces or scenes, typically generated by depth cameras or LiDAR sensors.
Policy (RL): A function (often a neural network) that maps observations to actions, representing an agent's behavior strategy in reinforcement learning.
PPO (Proximal Policy Optimization): A popular on-policy RL algorithm that updates policies via clipped objective function, balancing sample efficiency and stability. Default for Isaac Lab.
Publish-Subscribe: A messaging pattern where publishers send messages to topics and subscribers receive them, decoupling senders from receivers (core to ROS 2).
Q
Qdrant: A vector database optimized for similarity search, used for storing and retrieving embeddings in RAG (Retrieval-Augmented Generation) systems.
Quaternion: A 4-dimensional representation of 3D rotation (w, x, y, z) that avoids gimbal lock, commonly used in robotics and graphics.
R
RealSense D435i: Intel's stereo depth camera with RGB, dual infrared cameras, and IMU, providing aligned RGB-D data for 3D perception in robotics.
Reinforcement Learning (RL): A machine learning paradigm where an agent learns to maximize cumulative reward through trial-and-error interactions with an environment.
Reward Function: In RL, a function that assigns a scalar reward to each state-action pair, guiding the agent toward desired behaviors.
RGB-D: Image data combining RGB color information with depth (distance) for each pixel, enabling 3D scene understanding.
Robot Operating System (ROS): An open-source middleware framework for robot software development, providing communication, hardware abstraction, and libraries. ROS 2 is the latest version with improved real-time performance and security.
RViz: The standard 3D visualization tool for ROS, displaying robot models, sensor data (point clouds, images), and planning results.
S
SAM (Segment Anything Model): Meta's foundation model for image segmentation, trained on 1 billion masks, capable of zero-shot segmentation of any object with optional prompts.
SDF (Simulation Description Format): An XML format used by Gazebo to describe robot models, environments, and simulation parameters (alternative to URDF).
Semantic Segmentation: Classifying each pixel in an image into object categories (e.g., road, car, person), providing dense scene understanding.
Sensor Fusion: Combining data from multiple sensors (camera, LiDAR, IMU) to produce more accurate and robust estimates than any single sensor.
Service (ROS 2): A request-response communication pattern in ROS 2, where a client sends a request and waits for a server's reply (e.g., motion planning service).
Sim-to-Real Transfer: The process of deploying policies trained in simulation to real robots, addressing the reality gap through domain randomization and accurate modeling.
SLAM (Simultaneous Localization and Mapping): Building a map of an unknown environment while simultaneously tracking the robot's location within it.
T
Task Decomposition: Breaking down high-level goals ("clean the table") into sequences of low-level sub-tasks ("detect objects", "grasp", "place in bin"), typically done by LLMs.
Tensor API: Isaac Sim's direct GPU memory access interface for observations/actions, avoiding CPU-GPU transfers. All data stays on GPU as PyTorch tensors.
TF2 (Transform Library 2): ROS 2's library for managing coordinate frame transformations, maintaining a tree of spatial relationships (e.g., base_link → camera_link → gripper_link).
Topic (ROS 2): A named communication channel in ROS 2's publish-subscribe pattern, carrying messages of a specific type (e.g., /camera/image_raw carries sensor_msgs/Image).
Trajectory: A time-parameterized path specifying robot joint positions/velocities at each timestep, used for smooth motion execution.
U
Unity ML-Agents: Unity's machine learning framework enabling RL training in Unity game engine environments, popular for game AI and some robotics applications.
Universal Scene Description (USD): An open-source 3D scene file format (developed by Pixar) supporting hierarchical composition, time-sampled data, materials, and lighting. Used by Isaac Sim.
URDF (Unified Robot Description Format): See "H" section. XML format for robot descriptions in ROS.
V
Vectorized Environment: Running N identical RL environments in parallel on the GPU, with batch operations (obs, actions, rewards as tensors). Enables massive throughput in Isaac Lab.
Vision Transformer (ViT): Transformer architecture applied to images by dividing images into patches and processing them as sequences, enabling state-of-the-art performance (used in Depth Anything, CLIP, SAM).
Visuomotor Policy: A neural network that maps visual observations (camera images) directly to motor commands (joint positions), learned via imitation or reinforcement learning.
VLA (Vision-Language-Action): Robotics paradigm integrating computer vision (3D perception), language models (task planning), and action models (control) to enable natural language-commanded robot behaviors.
W
Waypoint: A discrete position along a planned path that a robot must reach, used in trajectory planning and navigation.
Whisper: OpenAI's speech recognition model trained on 680,000 hours of multilingual audio, achieving human-level transcription accuracy for voice-to-text conversion.
Workspace: In robotics, the reachable space for a robot's end-effector; also refers to a ROS 2 directory containing packages built with colcon.
X
X11: A windowing system for Unix-like operating systems, used to display GUI applications from Docker containers on the host display.
Y
YAML (YAML Ain't Markup Language): A human-readable data serialization format used extensively in ROS for configuration files, launch files, and parameter definitions.
YOLO (You Only Look Once): A family of real-time object detection models that process images in a single forward pass, popular for robotics vision but limited to pre-defined object classes.
Z
Zero-Shot Learning: A machine learning paradigm where a model performs tasks it wasn't explicitly trained on, by leveraging knowledge transfer (e.g., CLIP detecting novel objects via text descriptions).
Total Terms: 100+
Last Updated: 2025-12-05 Note: Terms are drawn from all chapters of the Physical AI & Humanoid Robotics textbook. For detailed explanations and context, refer to the specific chapters where each term is introduced.