Artificial intelligence
Gaussian belief propagation
Definition
Gaussian belief propagation performs probabilistic inference in a Gaussian factor graph by passing local information-form messages between variable and factor nodes. It can distribute estimation or planning computations without assembling one central dense system.
Also known as: GBP
Updated
Local messages solve a Gaussian graph
A factor graph separates unknown variables from factors that encode measurements, priors, or constraints. In Gaussian belief propagation, each message is a Gaussian information contribution. Variable nodes combine incoming information, while factor nodes eliminate their other connected variables and send a new local message.
Ortiz and colleagues present GBP as a distributed inference algorithm that uses local storage, local computation, and message-passing communication. On a tree-structured graph, belief propagation obtains exact marginals with forward and backward passes. Graphs with cycles use repeated loopy updates; the result and convergence then depend on the graph and numerical conditions.
Estimation and planning can share the machinery
Robotics factor graphs commonly represent sensor fusion, mapping, and pose estimation. GBP can update only the affected neighbourhood when a factor changes and can pass messages asynchronously across processors or robots. Its locality is attractive when no single processor should hold the entire problem.
The same inference view can encode trajectory variables and collision factors. A 2026 preprint by Jin and colleagues applies a distributed dynamic GBP variant to online multi-arm motion planning. The authors report dual-arm hardware experiments and simulation results, including their stated completion and solve-time figures. Those are results for DD-GBP and its evaluated scenes, not properties guaranteed by GBP in general.
A distributed solver still has global failure modes
Local messages do not make a difficult graph exact. Loops can cause slow convergence, oscillation, or overconfident marginal estimates. Linearisation is needed when robot models or collision costs are nonlinear, so a solution can depend on the current expansion point.
Message scheduling, damping, communication delay, and numerical conditioning affect practical performance. Distributed operation also does not remove coupled constraints: two arms can still collide if their shared factors or messages are missing, stale, or incorrectly modelled. GBP should be evaluated against a central solution or known ground truth on the graph structures and update rates that a robot will actually use.
Sources
Related terms
Factor graph
A factor graph is a bipartite graph whose variable nodes represent unknown quantities and whose factor nodes represent functions involving subsets of those variables. Robotics systems use this structure to combine local motion, sensor, and prior constraints in estimation and optimisation problems.
Pose graph optimization
Pose graph optimization estimates a globally consistent set of robot or sensor poses from noisy relative-pose constraints represented as edges in a graph. It is commonly used as a back-end calculation in simultaneous localization and mapping.
Sensor fusion
Sensor fusion combines information from multiple sensors or estimation sources to produce a shared estimate. The combination must account for coordinate frames, timing, uncertainty, and dependence between inputs.
Motion planning
Motion planning finds a robot movement from an initial state to a goal while satisfying constraints such as collision avoidance. A planner may produce a geometric path, a timed trajectory, or a sequence of controls.