Localization accuracy of a robot platform using indoor positioning methods in a realistic outdoor setting

Robotic developments can be classified according to their area of application in indoor and outdoor systems (Bechar and Vigneault, 2017). Depending on the application area, they are equipped with respective sensor systems. Independent of the actual environment, precise localization for a reliable navigation is of utmost importance in order to perform tasks autonomously.

Most agricultural robots, such as the field robots BoniRob (Göttinger et al., 2014) and Ladybird (Underwood et al., 2015) using an omnidirectional drive and wheels for rough terrain, are developed for outdoor applications. Global navigation satellite systems (GNSS) and real-time kinematic (RTK) correction provide the necessary absolute localization accuracy for such outdoor applications. In open space, such systems work very well; but shading effects, multipath effects, as well as poor network coverage in rural areas can lead to a severely reduced localization accuracy, rendering the system non-operational (Ehsani et al., 2003). Further, such sensor systems provide no functionality in an indoor environment. In contrast, robots for indoor applications, which are mostly encountered in industrial fields such as logistics, manufacturing, or horticulture (Spray Robot in Shamshiri et al., 2018), usually rely on artificial marks or rail systems for their operation. However, current trends, which strive for a complete automation of agricultural processes, for example, in indoor horticultural applications, require unrestricted navigation. Alternative absolute positioning methods, in particular, laser scanners and Monte Carlo localization, can fulfill this need by providing reliable indoor localization (Hertzberg, 2012; SPARC, 2015).

Previously published agricultural robotic developments mainly focus on engineering aspects while stressing the importance of an accurate positioning system for a reliable use (Bechar and Vigneault, 2016). Still, only rarely, these positioning systems are thoroughly evaluated.

Röwekämper et al. (2012) focused on the evaluation of the integrated localization and motion planning system of an industrial and mobile manipulation robot, that is, the KUKA omniRob system. The localization system is built upon standard components in robotics, namely, two laser scanners and Monte Carlo localization. This omni-drive steering platform is developed specifically for logistics in industrial processes and is equipped with Mecanum wheels. Therefore, operation is only possible on leveled surfaces.

In contrast, the in-house agricultural robot platform “Mathilda” is intended to serve as a carrier vehicle for a wide range of applications, and is therefore equipped with a differential drive system and pneumatic tires for motion in rough terrain. So far, indoor bench tests using a single laser scanner are performed to evaluate the positional accuracy of Mathilda under laboratory conditions (Supper et al., 2021). Agricultural robots with sufficient flexibility and a high degree of autonomy to navigate in a wide range of environments under various conditions depend highly on accurate localization (Shamshiri et al., 2018). Further, under harsh conditions or specific circumstances, outdoor positioning methods might not be available and indoor positioning methods need to take over in order to ensure reliable localization. Thus, the aim of this work is to evaluate the localization accuracy of the field robot Mathilda using indoor positioning methods under realistic outdoor conditions. For this purpose, Mathilda navigates to predefined target points in an outdoor environment and the actual position is independently determined using a motion capture system.

Materials and methods

2.1

Robot platform

The robot platform used here Mathilda (Figure 1a) has been already described in previously published papers (Supper et al., 2019, 2021). For completeness, the developed robot has a width of 850 mm, a length of 1100 mm, and a weight of 250 kg. The chassis consists of a welded steel construction. The HDC 2460 motor controller (Roboteq, Scottsdale, AZ, USA) controls the two front drive wheels, which are used to realize the steering via a speed control. These are designed as pneumatic tires with a diameter of 500 mm and a width of 210 mm. The rear axle consists of two swivel castors with 270 mm diameter and 85 mm width. The maximum possible driving speed in this configuration is 0.9 m/s. Sensors for navigation are a LIDAR scanner VLP16 (Velodyne, San Jose, CA, USA), Inertial Measurement Unit (IMU) Brick 2.0 (TinkerForge, Schloß Holte-Stukenbrock, Germany), and incremental encoders RVP510 (ifm electronic, Essen, Germany) on the drive wheels.

Experimental setup. (a) Picture of the field robot “Mathilda.” (b) Picture of the outdoor laboratory. (c) Map of the outdoor laboratory generated using a laser scanner and the ROS “gmapping” node with the starting area (yellow), the commonly used paths (blue), and target positions (green), as well as the Vicon Vantage V5 (red) camera positions indicated. ROS, robot operating system
Abbildung 1. Versuchsaufbau. a) Bild des Feldroboters “Mathilda”. b) Bild des Freilandlabors. c) Mit einem Laserscanner und dem ROS “gmapping”-Knoten erstellte Karte des Freilandlabors mit Angabe des Startbereichs (gelb), der gemeinsamen Wege (blau) und Zielpositionen (grün) sowie der Kamerapositionen der Vicon Vantage V5 (rot).

The robot software is based on the robot operating system (ROS) framework. For the operation of the robot, a software concept is used, which is based on the ROS navigation stack. This software calculates a path with the help of a global planner. A local planner takes care of the path following. The used “Elastic Bands” local planner determines the driving speed of the robot, which is limited by the adjustment of the local planner to 0.2 m/s and the angular speed to 0.15 rad. This software concept is complemented by in-house developments in robot kinematics and localization, which adapt the software to specific characteristics of the developed robot platform (Supper et al., 2019). For indoor localization, an “adaptive Monte Carlo localization algorithm” (AMCL) is used. This is an algorithm for robots to localize using a particle filter. With the help of a given map of the environment, the algorithm estimates the position and orientation of a robot as it moves and senses the environment (Fox et al., 1999).

2.2

Experimental design

All experiments were carried out in an outdoor laboratory of University of Natural Resources and Life Sciences, Vienna, which is a paved area (46 m²) including planting boxes and is surrounded by a fence (Figure 1b). In order to recognize the fence as a boundary with the laser scanner, a privacy screen was attached to the fence. The content of the plant boxes themselves was not part of the experiment. A spacing of 1.2 m between the planting boxes was chosen to ensure that the mobile robot platform can drive in between them but cannot turn around. This test setup is supposed to bring the robot’s navigation system close to its limit.

Only indoor navigation systems were used, as GNSS accuracy is low and not reliable in the urban environment. Point and orientation determination was carried out with the combination of a map of the outdoor laboratory and a laser scanner using rangefinder models of the ROS package “AMCL” (Mahtani et al., 2016). The map was created in advance by the laser scanner and the software of the ROS package “gmapping” by manually driving the mobile robot platform in the outdoor laboratory (map in Figure 1c). In the selected area, two target positions, including the point and the orientation, were defined and autonomously approached by the mobile robot Mathilda 15 times each. For these repetitions, the mobile robot platform was moved manually to a starting area. A successful run, defined by the ROS “target reached” notification, is included in the evaluation if the robot reaches the target position in certain limits, which are defined by the ROS parameters “xy_goal_tolerance” of 0.1 m and “yaw_goal_tolerance” of 0.1 rad (5.73°). The limits were adjusted to the robot width and the row distance between the plant pots. Runs were aborted when the safety system was triggered, for example, the robot hit a plant box. The motion analysis system Vantage V5 (Vicon Systems Ltd, Oxford, England) consisting of the Vicon Vantage cameras, a network switch, reflector balls, and the Vicon Tracker software was used as an external reference measurement system to determine the actual robot position. For this purpose, eight cameras were positioned in u-shape around the outdoor laboratory (Figure 1c). The cameras were then aligned and adjusted to the current lighting conditions. Further, the motion analysis system was calibrated according to the manufacturer’s protocol. The robot platform was covered with six reflective markers of 14 mm diameter. These formed an object that was tracked by the software Vicon Tracker both in terms of point and orientation.

2.3

Evaluation

Data analysis was carried out using in-house MatLab algorithms (MathWorks, Natick, MA, USA). The x and y axes are defined as a Cartesian coordinate system, with the x axis being orientated along the line of plant boxes (Figure 1c). Distance errors are defined as the difference between the actual point and the target point: (1) $Δ x = x - x_{0}$ \Delta x = x - {x_0} (2) $Δ y = y - y_{0}$ \Delta y = y - {y_0} with x₀ and y₀ being the coordinates of the target point and x and y being the coordinates of the actual finishing point. The absolute distance error d of the actual point and the target point is calculated using the Euclidean distance as follows: (3) $d = \sqrt{{(x - x_{0})}^{2} + {(y - y_{0})}^{2}}$ d = \sqrt {{{(x - {x_0})}^2} + {{(y - {y_0})}^2}}

Angular errors ΔΨ are defined as the difference between the actual finishing orientation Ψ and the target orientation Ψ₀. (4) $Δ Ψ = Ψ - Ψ_{0}$ \Delta \Psi = \Psi - {\Psi _0}

Results and discussion

For position 1, the positioning error along the x-axis Δx was 147.6 mm with a narrow distribution shown by a standard deviation of 30.2 mm (Figure 2a). In contrast, the positioning error along the y-axis Δy had a much lower mean of 2.1 mm and a similar standard deviation of ±29.7 mm (Figure 2b). In combination, these results led to a mean absolute distance error d of 150.7 ± 28.1 mm, mainly caused by exceeding the target point in x-direction (Figure 2c).

Boxplot of the deviation of the actual position to the target position. (a) Distance error in x-direction, (b) distance error in y-direction, (c) absolute distance error d, and (d) angular error ΔΨ for the two positions and the positions pooled together. Green lines indicate hysteresis threshold. Data are from 11 and 10 successful runs for position 1 and 2, respectively.
Abbildung 2. Boxplot der Abweichung der Ist-Position von der Soll-Position. a) Abstandsfehler in x-Richtung, b) Abstandsfehler in y-Richtung, c) absoluter Abstandsfehler d und d) Winkelfehler ΔΨ für die beiden Positionen und die zusammengefassten Positionen. Grüne Linien zeigen die Hysterese. Die Daten stammen aus 11 bzw. 10 erfolgreichen Läufen für Position 1 bzw. 2.

The evaluation of position 2 showed a different behavior with a higher mean absolute distance error d of 251.9 ± 47.2 mm (Figure 2a–c). The higher standard deviation of d in position 2 indicated a broader distribution in comparison to position 1. Interestingly, a detailed view of position 2 showed the different behavior compared to position 1 along the x- and y-axes with a distribution mean of 31.7 mm in x-direction and −240.5 mm in y-direction. This might be due to the fact that the local environment varies for each position and individual surface properties of the surroundings.

In general, no trend in the absolute distance error d could be observed for positions 1 and 2. Taken together, the mean absolute distance error d over all positions was 198.9 ± 63.9 mm. Figure 2c shows that all values exceed the given “xy_goal_tolerance” parameter. Therefore, deviations were caused by an incorrect position estimate and not by hysteresis.

The mean angular error ΔΨ over all positions was 4.9° ± 2.8°. The angular error ΔΨ showed a different behavior for positions 1 and 2. The mean angular error ΔΨ in position 1 was −2.5° ± 3.1° and in position 2, we observed a mean angular error ΔΨ of −7.0° ± 1.4°. While in position 1, more than 75% was within the “yaw_goal_tolerance”, in position 2, approximately 95% failed this given limit, which suggests an incorrect position estimate too (Figure 2d).

Of 30 planned test runs, 21 were performed successfully (Pos. 1: 11; Pos. 2: 10) while 9 of them ended prematurely. The reasons for this were collisions with the plant pots and triggering of the safety system. Whereas in x-direction the robot overshot the target, the distance error in y-direction could lead to a collision of the robot with the plant pots, and therefore to an incomplete run.

In comparison, Röwekämper et al. (2012) reported an average accuracy of 5 mm/0.15° in an indoor setting. This much lower positioning error could be explained by two main factors. First, the used Kuka robot platform was equipped with an omnidirectional drive and two laser scanners. In terms of fine positioning, this system offers advantages over a differential drive such as that used by the Mathilda field robot. The large pneumatic tires also had a disadvantage in terms of positioning accuracy. Second, Röwekämper et al. (2012) performed the experiments in an indoor setting, which provides clearly edges for detection in comparison to our outdoor laboratory setting.

Previous indoor tests with the field robot Mathilda showed an average positioning error of 45.1 mm/0.4° (Supper et al., 2021). In comparison to the used outdoor laboratory, Supper et al. (2021) described a clearly demarcated indoor test-bench without obstacles. This suggests that the decreased positional accuracy could be explained by the given outdoor scenario itself. The small space between the plant pots and the robot's differential drive led to accumulated errors in positioning, and eventually to the large positioning errors. Compared to the technically more complex omnidirectional drive systems, such as BoniRob, Ladybird, and the Kuka OmniRob system, Mathilda uses a simpler, implementable differential drive. Despite the straightforward designed driving system, reliable navigation should be possible. For differential drive systems, the non-holonomic constraint means that the robot cannot drive straight to a goal that is not in line with its orientation. It must either rotate to the desired orientation before moving forward or rotate as it moves. Taking into account the mechanical limitations of differential drive systems in path planning algorithms could result in better positioning performance.

Conclusion

The idea of the field robot Mathilda was to set up a cheap and lightweight field robot for a wide range of applications. Here, we evaluated the field robot Mathilda using a laser scanner and an indoor positioning method for navigation to predefined target points in an outdoor environment. Results show a mean absolute distance error d over all positions of 198.9 mm determined in the test runs in the presented outdoor laboratory. Most likely, limitations by the differential drive system are responsible for this unsatisfying result. To compensate for these disadvantages, path planning must be adapted to the differential drive, which requires further research. In the next step, positioning accuracy including obstacle detection needs to be evaluated in a dynamic environment.

eISSN:: 2719-5430
Language:: English

Publication timeframe:: 4 times per year
Journal Subjects:: Life Sciences, Ecology, other

Journal RSS Feed

Localization accuracy of a robot platform using indoor positioning methods in a realistic outdoor setting

Published Online: Jun 23, 2022

Page range: 133 - 139

Received: Sep 22, 2021

Accepted: Nov 18, 2021

DOI: https://doi.org/10.2478/boku-2021-0014

KeywordsRobotics, automation, Monte Carlo localization, precision agriculture, sensor technology

© 2021 Georg Supper et al., published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 3.0 License.

Figure 1

Figure 2

Keywords
Robotics, automation, Monte Carlo localization, precision agriculture, sensor technology