GitHub

POMDPs.jl Integration

GaussianFilters ships with a POMDPs.jl package extension that activates automatically when both packages are loaded. The extension wires AbstractFilter into the POMDPs belief-updater interface so that a Kalman, Extended Kalman, or Unscented Kalman filter can be used directly as a POMDPs.Updater — including with simulators such as HistoryRecorder and policy/planner code that expects the standard POMDPs interface.

Quick start

using GaussianFilters
using POMDPs
using POMDPTools

# 1. Build a filter as usual
dmodel = NonlinearDynamicsModel(f, W)
omodel = NonlinearObservationModel(h, V)
ekf    = ExtendedKalmanFilter(dmodel, omodel)

# 2. Wrap it as a POMDPs.Updater
updater = pomdps_updater(ekf)

# 3. Pass it to any POMDPs.jl simulator
hist = simulate(HistoryRecorder(max_steps=60), pomdp, policy, updater)

The wrapper is needed because POMDPs.jl simulators dispatch on the abstract type POMDPs.Updater, and AbstractFilter cannot subtype POMDPs.Updater directly without making POMDPs.jl a hard dependency.

Direct dispatch (without the wrapper)

For simpler use cases that don't need the full POMDPs.simulate machinery, the extension also overloads POMDPs.update and POMDPs.initialize_belief directly on AbstractFilter:

b1 = POMDPs.update(ekf, b0, action, observation)

This is convenient when integrating with code that calls POMDPs.update generically but does not require the <:Updater subtype constraint.

Initial beliefs from distributions

The extension supports initializing a GaussianBelief from any multivariate normal distribution. This is useful when the initial state of a POMDP is given as an MvNormal:

using Distributions
prior = MvNormal([0.0, 0.0], [1.0 0.0; 0.0 0.5])
b0    = POMDPs.initialize_belief(updater, prior)

API

GaussianFilters.pomdps_updater — Function
pomdps_updater(filter::AbstractFilter)

Wrap a GaussianFilters filter in a POMDPs.Updater so it can be passed to POMDPs.jl simulators like HistoryRecorder. Requires POMDPs.jl to be loaded; the implementation lives in a package extension.

source

Examples

Two example scripts live in the examples/ directory:

  • pomdps_integration.jl — a minimal demonstration of using a KalmanFilter through POMDPs.update.

  • pendulum_ekf_ilqr.jl — closed-loop stabilization of a noisy inverted pendulum observed only through its angular velocity. Defines a PendulumPOMDP, an ILQRPolicy <: POMDPs.Policy that runs iterative LQR on the belief mean (certainty equivalent control), wraps the EKF with pomdps_updater, and drives the whole closed loop through POMDPs.simulate.