Original author: Maxime Bouton, [email protected]

This repository consists of different driving scenarios formulated as POMDPs. It provides a generative model for computing policies. A few of them have explicit transition and observation models. This library is based on "POMDPs" v0.9 instead of "POMDP" v0.8 in original library.

To install this package, first add the SISL registry and the JuliaPOMDP registry such that all dependencies are automatically installed. You can run the following in the bash:

git clone https://github.com/AutomobilePOMDP/AutomotivePOMDPs.jl
cd AutomotivePOMDPs.jl
julia

pkg> add POMDPs
using POMDPs
POMDPs.add_registry()
pkg> registry add "https://github.com/sisl/Registry"
pkg> instantiate

Run docs/Urban POMDP tutoial.ipynb for a visualization of the simulation environment.

This package exports the following POMDP Models:

• SingleOCPOMDP: Occluded crosswalk with one single pedestrian. Discrete states and observations with explicit transition and observation model.
• SingleOIPOMDP: Occluded intersection with one single car. Discrete states and observations with explicit transition and observation model.
• OCPOMDP: Occluded crosswalk with a flow of pedestrian. Generative model implementation with continuous state and observations
• OIPOMDP: Occluded T intersection with a flow of cars driving in multiple lanes. Generative model implementation with continuous state and observations
• UrbanPOMDP : Occluded T intersection with crosswalks, flow of cars and pedestrians. Generative model implementation with continuous state and observations

These models are defined according to the POMDPs.jl interface. To see how they are parameterized, toggle the documentation using ? or use the function fieldnames if documentation is not yet written.

Currently, only the SingleOCPOMDP environment can be solved using online and offline methods, examples are as follows:

using POMDPs
using AutomotivePOMDPs
using POMDPPolicies
using POMDPSimulators
using BeliefUpdaters
using ParticleFilters
using AutomotiveDrivingModels
using AutomotiveSensors
using AutoViz
using Reel
using QMDP
using ARDESPOT

pomdp = SingleOCPOMDP(ΔT=0.5, max_acc=1.5, p_birth=1.0, action_cost=-0.2, γ=0.99)

qmdp_policy = solve(QMDPSolver(max_iterations=1000), pomdp)

function qmdp_upper_bound(pomdp, b)
    return value(qmdp_policy, b)
end

solver = DESPOTSolver(lambda=0., K=100, bounds=IndependentBounds(DefaultPolicyLB(RandomSolver()), qmdp_upper_bound, check_terminal=true),
default_action=SingleOCAction(-2.0), bounds_warnings=false)
planner = solve(solver, pomdp)

up = SingleOCUpdater(pomdp);

hr = RolloutSimulator(rng=rng, max_steps=50)

@show qmdp_reward = POMDPs.simulate(hr, pomdp, qmdp_policy, up);
@show despot_reward = POMDPs.simulate(hr, pomdp, planner, up);

• AutomotiveDrivingModels.jl
• POMDPs.jl