• 设为首页
  • 点击收藏
  • 手机版
    手机扫一扫访问
    迪恩网络手机版
  • 关注官方公众号
    微信扫一扫关注
    迪恩网络公众号

carlosmccosta/dynamic_robot_localization_tests: Tests for dynamic_robot_localiza ...

原作者: [db:作者] 来自: 网络 收藏 邀请

开源软件名称(OpenSource Name):

carlosmccosta/dynamic_robot_localization_tests

开源软件地址(OpenSource Url):

https://github.com/carlosmccosta/dynamic_robot_localization_tests

开源编程语言(OpenSource Language):

Shell 99.8%

开源软件介绍(OpenSource Introduction):

Dynamic Robot Localization tests

Overview

The dynamic_robot_localization_tests is a ROS package that aims to test the ROS dynamic_robot_localization package (or any other localization system that relies in laser / point cloud sensor data).

The main testing configurations are managed in launch/localization_tests.launch and are specialized for each environment / bag in launch/environments.

The localization_tests.launch file can be customized to change the localization system that is going to be tested.

They can use rosbags or live sensor data from the Gazebo simulator (rosbag play started with --pause --> press space bar to start publishing msgs).

The test results along with environment screenshots / videos are available in this shared folder

Testing platforms

The localization system was tested on laser sensor data retrieved from three different robots and executed on the same computer in order to allow direct comparison of computation time. This computer was a Clevo P370EM3 laptop with a Intel Core i7 3630QM CPU at 2.4GHz, 16 GB of RAM DDR3, NVidia GTX680M graphics card and a Samsung 840 Pro SSD.

The sensor data was recorded into rosbags, and is publicly available in the datasets folder.

The hardware specifications of the lasers used along with all the detailed results and experiments videos are available in this shared folder.

Jarvis platform

The Jarvis platform is an autonomous ground vehicle equipped with a SICK NAV 350 laser for self-localization (mounted about 2 meters from the floor) and a SICK S3000 laser for collision avoidance (mounted about 0.20 meters from the floor). It uses a tricycle locomotion system with two back wheels and a steerable wheel at the front. In Fig. 1 the robot is performing a delivery task with the package on top of a moving support. The 3 DoF ground truth was provided by the SICK NAV350 system and relied on 6 laser reflectors (with 9 cm of diameter) to perform the pose estimations (it is certified for robot docking operations with precision up to 4 millimeters).

Jarvis robot

Figure 1: Jarvis robot

Pioneer 3-DX

The Pioneer 3-DX shown in Fig. 2 is a small lightweight robot equipped with a SICK LMS-200 laser (mounted about 48 cm from the floor) and a Kinect (mounted about 78 cm from the floor). It uses a two-wheel two-motor differential drive locomotion system and can reach a linear speed of 1.2 m/s and angular velocity of 300º/s. The 3 DoF ground truth was provided by 8 Raptor-E cameras and according to RGB-D SLAM Dataset and Benchmark, it had less than 1 cm in translation error and less than 0.5 degrees in rotation error.

Jarvis robot

Figure 2: Pioneer 3-DX robot

Guardian platform

The Guardian platform is an autonomous mobile manipulator equipped with a Hokuyo URG-04LX laser in the front and a Hokuyo URG-04LX_UG01 laser in the back (both mounted about 0.37 meters from the ground). The front laser had a tilting platform which allows 3D mapping of the environment. The arm is a SCHUNK Powerball LWA 4P and in Fig. 3 it is attached to a stud welding machine (in simulation it is attached to a video projector). It uses a differential drive locomotion system and can be moved with wheels or with tracks. The 6 DoF ground truth for the Labiomep environment was provided by a Qualisys motion tracking system with 12 infrared cameras (tracking reflective markers). The 3 and 6 DoF ground truth for the simulated ship interior was provided by the Gazebo simulator.

Guardian robot

Figure 3: Guardian robot

Guardian robot simulated in Gazebo

Figure 4: Guardian robot simulated in Gazebo

Testing environments

The localization system was tested in different environments and used the Jarvis platform in a large room with a RoboCup field, the Pioneer 3-DX in a large industrial hall, the Guardian platform in indoor environments and a Kinect in a flying arena.

Jarvis in robocup field

The RoboCup field (shown in Fig. 5) occupies half of a large room (with 20.5 meters of length and 7.7 meters of depth). It has two doors, several small windows and two large glass openings into the hallway. Several tests were performed with the robot at speeds ranging from 5 cm/s to 50 cm/s in this environment and up to 2 m/s using the Stage simulator. These tests were performed with two different movement paths. The first is a simple rounded path that aimed to test the robot in the region of space that had better ground truth (due to its position in relation to the laser reflectors). The second path was more complex and contained several sub paths with different velocities and shapes. It was intended to evaluate the localization system with typical movements that mobile manipulators require, such as moving forward and backwards with or without angular velocity and stopping at the desired destination.

Launch files with synchronized ground truth available in this folder

Rosbags available in this folder.

Maps available in this folder.

JINT Jarvis dataset available in this folder.

Jarvis testing environment

Figure 5: Jarvis testing environment

Pioneer in industrial hall

The industrial hall is a large room with several tables and objects spread around. Four tests were performed with the Pioneer in this environment. The first was a 360º path with few objects in the middle of the room (shown in Fig. 6), while the remaining 3 tests were done with a lot of large objects that significantly reduced the field of view of the robot laser (as can be seen in Fig. 7).

The rosbags present in this folder only have laser data, and the associated launch files corrected the TF tree and synchronized the ground truth with the laser time stamps.

The original bag files are available at RGB-D SLAM Dataset and Benchmark (special thanks to the team who made this extensive dataset).

Maps available in this folder.

Industrial hall

Figure 6: Industrial hall

Industrial hall with objects in the middle

Figure 7: Industrial hall with objects in the middle

Kinect in flying arena

The flying arena shown in Fig. 8 is a large room in which several objects were added to test 6 DoF pose tracking. In these tests the Kinect was moved by the operator in three different paths. The first was a smooth fly movement over the testing scene, while the other two aimed to test paths with mainly translations and rotations. This environment had a ground truth provided by Vicon cameras and according to the authors of the dataset it had sub-centimeter accuracy.

Launch files available in this folder (requires the download of the kinect rosbags from the ASL datasets webpage to this folder).

Maps available in this folder.

Special thanks to the ASL team who made this dataset.

Flying arena environment

Figure 8: Flying arena environment

Guardian in Labiomep

The LABIOMEP environment is a large room with 10 meters of length and 15 meters of depth (panoramic image 1, panoramic image 2) equipped with 12 Oqus motion tracking cameras able to achieve sub-centimeter accuracy when tracking reflective markers.

This dataset was made to test internal and external 6 DoF pose tracking of the Guardian platform when climbing a ramp in 3 different environment setups with increasing complexity.

The first environment setup had very few objects close to the robot and aimed to test the robustness of the robot pose tracking systems when using sensor data with high noise.

In the second environment setup it was added several small cones around the ramp in order to test if the robot pose tracking accuracy improved.

In the last environment setup the cones were replaced with a L shaped wall with several vertical supports in order to test the accuracy of the tracking algorithms in a environment with a large amount of planar geometry (similar to a ship interior).

Launch files available in this folder.

Rosbags available in this folder.

Maps available in this folder.

Hardware used:

Overview of testing environment

Figure 9: Overview of testing environment

First testing environment with only the ramp

Figure 10: First testing environment with only the ramp

Second testing environment with cones around the ramp

Figure 11: Second testing environment with cones around the ramp

Second testing environment with robot climbing the ramp

Figure 12: Second testing environment with robot climbing the ramp

Third testing environment with ship interior walls

Figure 13: Third testing environment with ship interior walls

Third testing environment with robot climbing the ramp

Figure 14: Third testing environment with robot climbing the ramp

Mapping of the third testing environment with the Kinect

Figure 15: Mapping of the third testing environment with the Kinect

Front view of the Guardian robot

Figure 16: Front view of the Guardian robot

Back view of the Guardian robot

Figure 17: Back view of the Guardian robot

1 of the 4 Oqus 310+ cameras in LABIOMEP

Figure 18: 1 of the 4 Oqus 310+ cameras in LABIOMEP

1 of the 8 Oqus 400 cameras in LABIOMEP

Figure 19: 1 of the 8 Oqus 400 cameras in LABIOMEP

Guardian in ship interior

The structured environment simulated in Gazebo is a large room with 12.4 meters of length and 8.4 meters of depth. It has 4 doors, several small windows and the walls have small ledges at regular intervals (as can be seen in Fig. 20). Given that the Guardian mobile manipulator is expected to work on the walls of this environment, several tests were devised with a path following the lower and right wall of the environment. The first test was done in a static environment clear of unknown objects and was meant to evaluate the best precision that the localization system could achieve. The second test was done in a cluttered environment and was designed to test the robustness of the localization system against static unknown objects, that were placed in the middle of the environment and close to the walls (to block sensor data from reaching known positions and analyze the robustness of matching unknown points that are close to known areas). The last test added a moving car to the scene and aimed to assess the impact of dynamic objects on the point cloud registration algorithms.

Launch files available in this folder.

Rosbags available in this folder.

Maps available in this folder.

Guardian testing environment

Figure 20: Guardian testing environment

List of related git repositories:

More info




鲜花

握手

雷人

路过

鸡蛋
该文章已有0人参与评论

请发表评论

全部评论

专题导读
热门推荐
阅读排行榜

扫描微信二维码

查看手机版网站

随时了解更新最新资讯

139-2527-9053

在线客服(服务时间 9:00~18:00)

在线QQ客服
地址:深圳市南山区西丽大学城创智工业园
电邮:jeky_zhao#qq.com
移动电话:139-2527-9053

Powered by 互联科技 X3.4© 2001-2213 极客世界.|Sitemap