Intention-Aware Control Facilitated by Polynomial Chaos Expansion

Author: Zengjie Zhang (z.zhang3@tue.nl)

A benchmark of intention-aware control of autonomous vehicles using polynomial chaos expansion (PCE).

Release

It is suggested to download the release branch or the releases for a cleaned up version.

Introduction

When an autonomous vehicle interacts with other traffic participants, the awareness of the intentions of these participants strongly impacts the safety and reliability of its decision-making. Fig. 1 shows an intersection case where an ego vehicle (EV) plans a left turn while avoiding collisions with opponent vehicles (OVs) and pedestrians from the opposite direction. Understanding the intentions of these participants (e.g., moving fast or slowly) facilitates the design of a safe and flexible policy.

Figure 1. An intersection scenario.

Approach

This project provides a correct-by-design solution for intention-aware control of uncertain autonomous agents. By formulating unknown intentions of traffic participants as discrete-valued random variables, we derive a stochastic control problem for epistemically uncertain agents with a belief-space specification. In this way, the risk is quantified as the probability of not satisfying certain belief-space specifications, and the control problem is aimed to restrict the risk with a predefined threshold. Then, we use stochastic expansion to convert this challenging stochastic control problem into a tractable deterministic control problem with an STL specification. This allows us to solve a challenging intention-aware control problem using off-the-shelf tools, such as shrinking-horizon MPC based on Mixed-Integer Planning (MIP). With assumed intention models, the solved intention-aware controller allows automatic policy adjustment and ensures strict risk limitation.

Fig. 2 shows the comparison between a conventional no-aware controller and the proposed intent-aware controller. The blue blocks are sampled OVs subject to a multi-modal distribution, corresponding to speeding up, constant speed, and slow down intents. The results clearly show that the EV with a no-aware controller may collide with a speeding up OV, while the intent-aware vehicle can effectively avoid collision with all possible OVs.

Figure 2. The comparison between no-aware (left) and intent-aware (right) controllers.

Installation

Requirements

Operating system

• Windows (compatible in general, succeed on 11)
• Linux (compatible in general, succeed on 20.04)
• MacOS (compatible in general, succeed on 13.4.1)

Python Environment

• Python 3.11
• Required Packages: numpy, numpoly, scipy, treelib, matplotlib, importlib-metadata.

Required Libraries

• gurobipy solver (license required, see How to Get a Gurobi License)
• stlpy toolbox (see Documentation or GitHub repository)
• chaospy toolbox (see Documentation or GitHub repository)

Quick Installation

1. Install conda following this instruction;

2. Open the conda shell, and create an independent project environment;

conda create --name intentaware python=3.11

3. In the same shell, activate the created environment

conda activate intentaware

4. In the same shell, within the intentaware environment, install the dependencies one by one

conda install -c anaconda numpy
conda install -c conda-forge numpoly
conda install -c anaconda scipy
conda install -c conda-forge treelib
conda install -c conda-forge matplotlib
conda install -c conda-forge importlib_metadata

5. In the same shell, within the intentaware environment, install the libraries gurobipy, stlpy, and chaospy:

python -m pip install gurobipy
pip install stlpy
pip install chaospy

6. Last but not least, activate the gurobi license (See How To). Note that this project is compatible with gurobi Released version 11.0.1. Keep your gurobi updated in case of incompatibility.

Quick Run

• Lead to the example directory;
• Lead to the overtaking or the intersection folder;
• Run the main script main.py;
• Rendering constraints may take up to 1 to 2 minutes, depending on the computational performance;
• Plotted figures automatically saved in the data subfolder.

Special visualization functions

If using config/record function, be sure to install ffmpeg first and make sure its installation folder shows up in the PATH environmental variable.

License

This project is with a BSD-3 license, refer to LICENSE for details.