Fisher Information Field: an Efficient and Differentiable Map for Perception-aware Planning

📌 For an earlier version of this work that was published at ICRA 19, please checkout the icra19 branch.

This repository contains an implementation of the Fisher Information Field (FIF for short), a map representation designed for perception-aware planning. The core function of the map is to evaluate the visual localization quality at a given 6 DoF pose in a known environment (consisting of 3D landmarks to localize against). It can be used with different motion planning algorithms (e.g., RRT-star, trajectory optimization) to take localization quality into consideration, in addition to common planning objectives (collision free, low dynamic cost, start and end states etc). The main advantage of FIF is its efficiency at the planning time: it is 1~2 order-of-magnitude faster than using the landmarks directly in our experiments. In addition, FIF is also differentiable, making it suitable to be used in gradient-based optimization.

Please read the preprint and watch the video for details. If you use this code in academic context, please cite the following papers:

• Zichao Zhang, Davide Scaramuzza. Fisher Information Field: an Efficient and Differentiable Map for Perception-aware Planning. arXiv preprint, 2020. PDF Video Bibtex
• Zichao Zhang, Davide Scaramuzza. Beyond Point Clouds: Fisher Information Field for Active Visual Localization. ICRA, 2019. PDF Video Bibtex

1. Credit
2. Packages
3. Install
4. Get Started

Credit

The implementation of the information field uses the voxel hashing algorithm implemented in Voxblox. The code we extracted from Voxblox repository is put in separate folders (voxblox and voxblox_ros), and the license is retained in third_party/LICENSE_voxblox. If you use the code specific to Voxblox, please see their repository and cite relevant publications accordingly.

The unreal_cv package uses some code from ESIM and UnrealCV for the interaction with UnrealEngine. If you use this part of the code, please check their repositories and refer to relevant publications accordingly.

Packages

The FIF and the integration with ROS and motion planning algorithms are implemented as the following ROS packages:

• act_map: the core functionality of the FIF, independent of ROS
• act_map_ros: ROS wrappers, including parameter loading, essential publishing and subscribing topics and launch file examples.
• act_map_msg: simple ROS messages
• act_map_exp: example integration of FIF with different motion planning algorithms

In addition, there are several third-party packages in third_party folder for convenience. Please see the README there for details. We also provide a ROS package unrealcv_bridge to interact with UnrealEngine for photorealistic rendering.

Install

The code is tested on Ubuntu 18.04 with ROS Melodic. It should be possible to use in other ROS/Ubuntu versions.

Install dependencies

First install ROS according to the official instruction. Then install some system dependencies:

sudo apt install python-pip python-catkin-tools python-vcstool libgoogle-glog-dev libatlas-base-dev libeigen3-dev  libsuitesparse-dev
pip install pyquaternion plotly GPy tqdm jupyter

Install Ceres from source if you have not yet according to the official instruction (you can directly start from the point of extracting the source code).

Install OMPL from source if you have not yet according to the official instruction.

latex is required for some plotting scripts but not the functionalities of the code.

Clone and Compile

In where you would like to put the workspace, execute the following

mkdir FIF_ws && cd FIF_ws
catkin config --init --mkdirs --extend /opt/ros/melodic   --cmake-args -DCMAKE_BUILD_TYPE=Release
cd src
git clone [email protected]:uzh-rpg/rpg_information_field.git
vcs-import < ./rpg_information_field/dependencies.yaml
touch minkindr/minkindr_python/CATKIN_IGNORE
cd ..
catkin build 

Download the Simulation Environment

We provide the photo-realistic simulation environment used in our paper as a standalone application, downloadable here. It is built with UnrealEngine (NVIDIA Isaac Sim 1.2) and uses UnrealCV to to query images and depths. Note that you will need the environment only for the motion planning experiments in act_map_exp but not the core functionalities of FIF.

To start the environment, run the following:

./IsaacSimProject.sh -WINDOWED

and you should see a first person view in a warehouse environment, where you can control with your keyboard and mouse like FPS games (press '`' to get mouse from the game). If you want to change the intrinsics of the camera (resolution and the FoV), you need to change the configuration in LinuxNoEditor/IsaacSimProject/Binaris/Linux/unrealcv.ini before starting the simulation.

We mostly use the simulation as a server: you can start it and let it run in the background, and we will query images and depths from it using the commands provided by UnrealCV (implemented in unrealcv_bridge package).

If you want to modify the environment or create your own, please see the instruction of how to set up and use the Isaac Sim. Note that this would require a relatively powerful computer.

Get Started

For the core function of the FIF, please check the documentation in the act_map package, including:

• An overview of the implementation as well as detailed description of some core concepts.
• How to extend the FIF to include new functionalities
• Runnable code for the simulation experiments in the paper

For the integration of FIF with motion planning algorithms, please check the documentation in the act_map_exp package, including

• Complete examples of using FIF in RRT* and trajectory optimization along with common planning objectives.
• Runnable code for the motion planning experiments in the paper
• Quantitative evaluation of the localization accuracy in photorealistic simulation.