Commit c4a194de authored by Guilhem Saurel's avatar Guilhem Saurel
Browse files

cosyslam v2

parent d1b6a475
Pipeline #17822 passed with stage
in 8 seconds
title: "CosySLAM: tracking contact features using visual-inertial object-level SLAM for locomotion"
subtitle: Submitted to 2022 IEEE ICRA - International Conference on Robotics and Automation
title: "CosySlam: investigating object-level SLAM for detecting locomotion surfaces"
subtitle: Submitted to 2022 IEEE/RSJ IROS - International Conference on Intelligent Robots and Systems
- César Debeunne ^1^
- Médéric Fourmy ^2,3^
- Yann Labbé ^5^
- César Debeunne ^1,2^
- Médéric Fourmy ^2^
- Yann Labbé ^3^
- Pierre-Alexandre Léziart ^2^
- Guilhem Saurel ^2^
- Joan Solà ^4^
- <a href="">Nicolas Mansard</a> ^1,2^
- Joan Solà ^2,4^
- <a href="">Nicolas Mansard</a> ^2,5^
- ^1^ ISAE-Supaero, Toulouse
- ^2^ LAAS-CNRS, Université de Toulouse
- ^3^ Artificial and Natural Intelligence Toulouse Institute, Toulouse
- ^3^ Inria, École normale supérieure, CNRS, PSL Research University, Paris
- ^4^ Intitut de Robòtica i Informàtica Industrial, Barcelona
- ^5^ Inria, École normale supérieure, CNRS, PSL Research University, Paris
- ^5^ Artificial and Natural Intelligence Toulouse Institute, Toulouse
## Abstract
A legged robot is equipped with several sensors observing different classes of information, in order to provide various estimates on its states and its environment.
While state estimation and mapping in this domain have traditionally been investigated through multiple local filters, recent progresses have been made toward tightly-coupled estimation.
Multiple observations are then merged into an a-posteriori maximum estimating several quantities that otherwise were separately estimated.
With this paper, our goal is to move one step further, by leveraging on object-based simultaneous localization and mapping.
We use an object pose estimator to localize the relative placement of the robot with respect to large elements of the environments, e.g. stair steps.
These measurements are merged with other typical observations of legged robots, e.g. inertial measurements, to provide an estimation of the robot state (position, orientation and velocity of the basis) along with an accurate estimation of the environment pieces.
It then provides a consistent estimation of these two quantities, which is an important property as both would be needed to control the robot locomotion.
We provide a complete implementation of this idea with the object tracker CosyPose, which we trained on our environment and for which we provide a covariance model, and with the SLAM engine Wolf used as a visual-inertial estimator on the quadruped robot Solo.
While blindfolded legged locomotion has demonstrated impressive capabilities in the last few years, further progresses
are expected from using exteroceptive perception to better adapt the robot behavior to the available surfaces of
contact. In this paper, we investigate whether mono cameras are suitable sensors for that aim. We propose to rely on
object-level SLAM, fusing RGB images and inertial measurements, to simultaneously estimate the robot balance state
(orientation in the gravity field and velocity), the robot position, and the location of candidate contact surfaces. We
used CosyPose, a learning-based object pose estimator for which we propose an empirical uncertainty model, as the sole
front-end of our visual inertial SLAM. We then combine it with inertial measurements which ideally complete the system
observability, although extending the proposed approach would be straightforward (e.g. kinematic information about the
contact, or a feature based visual front end). We demonstrate the interest of object-based SLAM on several locomotion
sequences, by some absolute metrics and in comparison with other mono SLAM.
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment