Skip to content

Latest commit

 

History

History
98 lines (70 loc) · 5.13 KB

README.md

File metadata and controls

98 lines (70 loc) · 5.13 KB
Raw

Vision-based Control of a Quadrotor in User Proximity: Mediated vs End-to-End Learning Approaches

Dario Mantegazza, Jérôme Guzzi, Luca M. Gambardella and Alessandro Giusti

Dalle Molle Institute for Artificial Intelligence (IDSIA), USI-SUPSI, Lugano, Switzerland

Proceedings of ICRA 2019

Abstract

We consider the task of controlling a quadrotor to hover in front of a freely moving user, using input data from an onboard camera. On this specific task we compare two widespread learning paradigms: a mediated approach, which learns an high-level state from the input and then uses it for de- riving control signals; and an end-to-end approach, which skips high-level state estimation altogether. We show that despite their fundamental difference, both approaches yield equivalent performance on this task. We finally qualitatively analyze the behavior of a quadrotor implementing such approaches.

Paper Info

Arxiv and related BibTeX

Dataset

The Dataset used is composed of 21 different rosbag files.

Each rosbag correspond to a single recording session. For the recording sessions we used software developed in house (available here). Each recording session as been manually trimmed to remove system start-up / takeoff / landing phases; this information is available in the gloabl_parameters.py script as bag_start_cut and bag_end_cut dictionaries with the bag name as key.

In each file we recorded multiple topics; for this paper we use the following topics:

Topic Description
/bebop/image_raw/compressed Drone's front facing camera feed
/optitrack/head Motion Capture information about the 6DOF user's head's pose. In sync with OptiTrack system timestamp
/optitrack/bebop Motion Capture information about the 6DOF drone's pose. In sync with OptiTrack system timestamp
/bebop/mocap_odom Motion Capture information about the 6DOF drone's pose + twist. In sync with Drone Arena timestamp
/bebop/odom Drone's Optical Flow odometry

In our test we randomly divided the whole dataset in train and test set as follows

Train set bagfiles Test set bagfiles
1, 2, 5, 7, 8, 9, 10, 11, 12, 13, 14, 16, 17, 18, 19, 20, 21 3, 4, 6, 15

The whole dataset can be downloaded (6.6 GB) here.

A Jupyter notebook implementing dataset extraction from rosbag files can be found here.

Code

The code is structured as follows.

.
├── script                  # Script directory
├── bagfiles                # Bagfiles main directory
│   ├── train               # Directory for the bagfiles selected for the train set
│   └── test                # Directory for the bagfiles selected for the test set
└── dataset
    ├── version1            # Model 1 dataset .pickle files
    ├── version2            # Model 2 dataset .pickle files
    └── version3            # Model 3 dataset .pickle files

The executable scripts are:

  • dataset_generator.py
  • keras_train.py

dataset_generator.py

Create the dataset files used by the models. After launching the script you will be prompted with a menu in order to select the type of dataset to create. Each model has its own dataset.

keras_train.py

Uses models (one or all at the same time) for prediction.

Figure: A representation of the three models. Model 1 (left), model 2 (center) and model 3(right).

All scripts are available here.

Video

The video accepted at ICRA 2019 is available here.

Learning Vision-Based Quadrotor Control in User Proximity

Dario Mantegazza, Jérôme Guzzi, Luca M. Gambardella and Alessandro Giusti

The video accepted at HRI 2019 is available here and related BibTeX here.

The relative github page is here.

Other videos are available here.

Errata

Fixed in v2 (final version for ICRA): In the paper submission for ICRA2019, each image in Fig.2 have the left and bottom plot with inverted axis. Also in the same figure the smaller plot is rotated by 90° to the right.