Skip to content

📚 Glossary & Key Concepts (Alphabetical)

This glossary lists every DSG‑JIT term in strict alphabetical order for instant lookup.


A

Agent

A time‑varying entity (robot/person) with a pose trajectory over time.

Agent Pose Node

A pose node associated with a specific agent + timestamp in the DynamicSceneGraph.


B

Bearing

A unit‑direction vector from a sensor to a target (camera or LiDAR).


C

Calibration (Extrinsics)

Rigid transform from a sensor frame to the robot/base frame.

Camera Intrinsics

3×3 calibration matrix mapping pixels → normalized rays.

CameraMeasurement

Typed camera measurement containing rays, image, timestamps, etc.

Chain Residual (SE3 Chain)

Residual enforcing that a sequence of SE(3) poses matches integrated motion.

Conversion Layer (Sensor Conversion Layer)

Converts raw sensor samples into typed measurements and measurement factors.


D

DSG — Dynamic Scene Graph

A time‑varying scene graph combining geometry + semantics + temporal structure.

DSG Layer

A semantic layer (rooms, places, objects, agents, voxels, etc).

DSGTrainer

Bi‑level optimization utility (inner GN, outer hyperparameter learning).

DynamicSceneGraph

Holds time‑indexed agent poses & temporal edges.

Dynamic Voxel Field

A voxel grid that changes over time (dynamic/deformable environments).

Dataset Stream

Any sensor input driven by FunctionStream or FileRangeStream.


E

Edge (Scene Graph Edge)

Relationship between two scene‑graph nodes (room–place, place–object, etc).

Euclidean Variable

A variable in ℝⁿ (e.g., landmark, voxel, calibration scalar).


F

Factor

A residual constraint relating one or more variables.

Factor Graph

Bipartite graph of variables ↔ factors.

FileRangeStream

Sensor stream reading structured samples from a text/log file.

FunctionStream

Sensor stream driven by a Python generator function.


G

Gauss–Newton Optimization

Nonlinear least‑squares solver implemented on manifolds in DSG‑JIT.

Graph Export (SceneGraph Export)

Converts graph nodes into VisNode/VisEdge for visualization.


H

Hybrid Factor Graph

A factor graph mixing SE(3), Euclidean, and voxel variables in one optimization.


I

IMU Delta Integration (Integrate IMU Delta)

Creates a small SE(3) increment from raw IMU samples.

Inner Optimization (Inner GN / Inner Gradient Descent)

Solves the factor graph while outer parameters are fixed.


J

Jacobian

Matrix of partial derivatives of residuals wrt variable components.

JIT Compilation

JAX compilation of residual functions for high‑performance SLAM.


L

Landmark

A 3D point in the world (usually Euclidean).

Landmark Prior

Unary factor constraining a landmark near a specific 3D location.

Layered Visualization

Rendering style placing node types on separate z‑planes.

LiDAR Scan

List of ranges/angles or rays converted to LidarMeasurement.

Lie Group

Differentiable group structure for SE(3), SO(3), etc.


M

Manifold

Non‑Euclidean space (like SE(3)) requiring special update rules.

Manifold Gauss–Newton

Lie‑group‑aware Gauss–Newton with retract/exp/log operations.

Manifold Metadata

Describes each variable’s manifold type + slice in packed state vector.

Measurement

Raw or typed sensor output (camera, lidar, imu, odom).

MeasurementFactor

Factor generated from a typed measurement.


N

Named Object

Scene‑graph object with semantic identifier (e.g. “chair_1”).


O

Object Node

A semantic object (chair/table/etc) connected to a place.

Observation Model

Predicts expected sensor measurement given state x.

Odometry

Relative motion factor between two poses.

Odom_tx

1D translation‑only odom helper for quick SE(3) edges.

Outer Optimization

Top‑level hyperparameter update loop above the inner GN solver.


P

Place

A mid‑level scene‑graph node (semantic waypoint, corridor, anchor).

Place3D

3D spatial place node used in realistic indoor environments.

Place Attachment

Semantic constraint linking an agent pose → place.

Plot Functions

Graph/scene‑graph visualization utilities in world.visualization.

Pose

An SE(3) rigid transform.


R

Range

Scalar distance reading (LiDAR, UWB, depth).

Range Observation

Factor derived from a distance measurement.

Residual

Prediction error: r(x) = f(x) − z.


S

Scene Graph

Hierarchical structure representing geometry + semantics.

SceneGraphWorld

High‑level API wrapping factor graph, semantic nodes, and helpers.

SE(3) — Special Euclidean Group

Rigid‑body transform group for robot poses.

SO(3)

Rotation‑only Lie group.

Semantic Edge

High‑level symbolic edge between semantic nodes.

SensorFusionManager

Polls sensors and delivers typed measurements + callbacks.

Sensor Streams (FunctionStream / FileRangeStream)

Abstract data sources for camera/lidar/imu streams.


T

Temporal Factor

Factor connecting time‑indexed nodes.

Temporal Layer

Dynamic layer storing agent trajectories and time relations.

Trajectory (Agent Trajectory)

Ordered sequence of time‑indexed SE(3) poses.

Type Weights (Learnable Weights)

Learned weights for residual types in bi‑level training.


V

Variable

Optimizable unknown in the factor graph.

VisNode / VisEdge

Lightweight structures for 2D/3D visualization.

Voxel

Volumetric grid cell.

Voxel Grid

3D grid of voxels for dense mapping experiments.

Voxel Observation Model

Residual connecting voxel center to observed 3D point.

Voxel Smoothness Residual

Factor connecting adjacent voxels enforcing smooth geometry.


W

World Model

Top‑level container for factor graph, variables, and optimization.


Z

Zero‑Mean Gaussian Noise

Standard additive noise assumption for measurements.