This paper introduces the Agilica Geo-Location (AGL) system, an Ultra-wideband (UWB) technologybased solution designed for precise positioning and tracking in challenging edge spaces. In these environments, traditional Global Navigation Satellite System (GNSS) coverage is often obstructed or unavailable, leading to inadequate positioning accuracy. Such scenarios are critical, especially for applications like autonomous drone landing on sea vessels, where onboard structures can obstruct a large part of the sky during the landing phase. Unlike many other UWB-based solutions, AGL supports both network-centric asset tracking and device-centric navigation simultaneously, enabling scalability and adaptability. The system's auto-calibration mechanism and robust signal processing ensure reliable performance even in challenging RF channel environments. AGL ofers decimeter-level accuracy, making it suitable for robotics, unmanned vehicles, and asset tracking. Key aspects of the AGL system implementation are presented, emphasizing the importance of tight time-synchronization and addressing the impact of non-line-of-sight (NLOS) propagation. The main goal of this study is to validate the AGL system's performance, showcasing its eficacy through benchmarking studies. The results demonstrate the AGL system capability in delivering decimeter-level accuracy, even in dynamic and high multipath scenarios.
CEUR ceur-ws.org
For positioning and tracking, many applications rely on the Global Navigation Satellite System (GNSS), commonly known as GPS, which on its own is a truly global and scalable system and works very well in the open spaces. However GNSS has limitations when it comes to operating in the edge spaces, such as inside buildings, tunnels, mines, onboard sea vessels oil rigs and wind-mil platforms, where the coverage may be unavailable, shaded, or the accuracy inadequate to support target applications. In this edge area, positioning accuracy down to couple of decimeters or better is needed, e.g., tracking and navigation of robotic devices in a warehouse, factory floor and autonomous take-of and landing of drones in a safe, reliable and eficient way from a ship in highly dynamic environment and challenging weather conditions. This provides motivation to invest research eforts to develop and realize innovative solutions that can work in areas with no, shaded or intermittent availability of GNSS signals, and thereby Proceedings of the Work-in-Progress Papers at the 13th International Conference on Indoor Positioning and Indoor Navigation (IPIN-WiP 2023), September 25 - 28, 2023, Nuremberg, Germany CEUR Workshop Proceedings extend and expand the economic and societal benefits of the GPS like positioning service to the edge environments. The Agilica Geo-Location system (AGL) targets this kind of need.
The AGL is based on the Ultra-wideband (UWB) wireless access technology. The AGL can deliver accurate positioning down to 10 cm and is well-suited for robotics, unmanned vehicles in air or on ground and for asset tracking. The UWB technology uses low power and wide bandwidth signals that allow them to capture transmit and receive timing of the signals with ifne resolution. The signals are robust to noise and interference. Based on the signal timing data and using a network of UWB anchor nodes and tags, positioning and tracking solutions can be developed for complex edge environments where other positioning technologies may struggle.
Compared to other navigation solutions such as based on ultrasonic sensors [
There are numerous UWB based positioning solutions proposed in the literature, but they
have gaps in terms of robustness, scalability, flexibility and evaluation in realistic scenarios
[
The AGL system ofers a unified positioning system and methods that support both
networkcentric asset tracking and device-centric navigation applications simultaneously, all powered
by the same anchor infrastructure [
The AGL technology garnered recognition, leading to a patent application filed in January
2020 [
This paper presents the AGL system and its key elements, with a particular focus on its efectiveness in real-world scenarios. To achieve the highly accurate positioning, tight time synchronization among anchors is crucial. The synchronization technique is presented in detail, along with an online mechanism to calibrate internal processing and antenna delays. The device-centric positioning engine, which relies on pseudorange measurements to the anchors, is also discussed, with a focus on identifying and mitigating various sources of error.
The main focus of this paper is to present and discuss the efectiveness of the AGL system with results from benchmarking studies conducted in real environments. Results of a synchronization scheme evaluated over a 4-hop anchor topology shows synchronization error under 0.2 ns measured in terms of the error standard deviation. This illustrates the potential spatial scalability that the AGL system can ofer. The positioning accuracy is evaluated in static and kinematic scenarios. In the first scenario, the tag is static while in the second scenario the tag is mobile with average speed between 4 km/h to 7 km/h and maximum speed between 7 km/h and 11 km/h. In the static scenario, the median position error is around 10 cm, which increased to almost 15 cm in the dynamic scenario. The error values for the 95th percentile point are 15 cm and 25 cm, respectively for the static and dynamic scenarios.
As shown in Fig. 1, the AGL system comprises a number of spatially distributed anchor or beacon nodes deployed at known positions. The anchors are connected to a backend-server through Ethernet or WLAN link, and from the server the anchors can be configured and their performance monitored. The system supports positioning and tracking of two type of tag devices over the UWB channel. First type of tag is for device-centric positioning and navigation, and is attached to client-end device, such as a robot, drone, or human, that needs to know its position for navigation. The second type of tag is for network-centric track and trace application, and is attached to dumb assets for which a backend server or a management system needs to know their position, but not the asset itself.
In the device-centric positioning, an unlimited number of such tags can be simultaneously localized, with the position update rate independent of the tag density, just like the GPS receivers. However the AGL is not a replica of GPS system. There are some fundamental diferences. For instance, the GPS relies on atomic clocks in the satellites, whereas the AGL is using consumer grade crystal oscillators with stability in 10 to 20 ppm range. Each AGL anchor node periodically broadcasts navigation messages over the UWB channel. The anchors access the UWB channel using a TDMA-based medium access control. The structure of the TDMA frame can be configured through the backend server. The anchors and tags use these messages to synchronize their time. The accurate time synchronization is crucial to realize positioning of the tag devices. Synchronization accuracy within a few tens of micro-seconds is acceptable for the channel medium access control but for positioning of the tags, synchronization within a nanosecond or sub-nanosecond range must be achieved and maintained. As a reference, a nanosecond error in time synchronization corresponds to 30 cm error in distance. Thus highly accurate time synchronization scheme is needed for synchronizing the anchor network. The AGL system implements a multi-hop synchronization scheme over the UWB wireless channel. The multi-hop is needed to provide spatial scalability because the UWB is a lowpower technology and the reliable communication range is quite limited. The ensuing Section 3 presents the details of this synchronization along with experimental results. Afterwards positioning method and performance results are presented in Section 4.
The common notion of time among the anchors can be achieved by sharing clock from a reference node. Due to limited communication range, clock has to be shared through sync hierarchy comprising a tiered structure of reference nodes. We assume that each node has knowledge of its corresponding master node. In the ensuing discussion it sufices to focus on the interaction between a master–slave duo. The time synchronization scheme in the AGL system has two components which are implemented in a sequential and recursive way: (i) estimation and tracking of the clock parameters; (ii) estimation and compensation for the internal propagation delays on the sender and receiver side.
For clock parameter estimation, one of the schemes implemented in the AGL system is described
in this section. Figure 2 illustrates the sharing of a clock between two nodes by exchanging
time-stamped packets. The master node, denoted by , serves as the reference node, while the
node sharing its clock is referred as the slave node, denoted by . Let and denote the time
measured by the local clock at and , respectively. As depicted in Fig. 2, the master node
periodically transmits packets at a specific interval. In the th packet, the master node includes
the transmit time-stamp of the packet, while the corresponding receive time-stamp is noted
at the slave node. Based on practical clock characteristics such as phase diference denoted by ,
clock skew denoted by , and frequency drift between the two clocks denoted by , a precise
relationship between the master and slave times can be modeled, as described in [
1 + 2 2 + .
(1) In equation (1), the observation noise afecting the measurement is represented by . It is assumed that either or is adjusted for the time-of-flight (ToF) of the packet, which can be calculated from the known positions of the anchors.
The clock parameters , , and are known to vary over time due to the efects of temperature,
component aging, vibration, and radiation on the clock circuitry [
(b) Outline of the PI control loop. within this window. Using these measurements, an estimate of the clock parameters can be obtained as =̂ ( H H)−1H y, C = 2(H H)−1. (2) Here C denotes the error covariance of the estimator and ≜ [ ] , y ≜ ⎢⎢ −+2 ⎡ ⎢ ⎣ −+1 ⋮ ⎤ ⎥ ⎥ , ⎥ ⎦
H ≜ ⎢ 1 −+2
⎡ 1 −+1
⎢
⎢
⎢⎢ ⋮
⎢
The estimator in (2) assumes that the noise is zero mean with covariance C = 2I. The
probability density function of the noise is otherwise unknown. In the absence of any knowledge
about the observation noise, it can be shown that the above estimator is optimal in the
leastsquares sense [
In the context of time synchronization for the positioning and navigation systems, estimating and compensating clock parameters is not enough when it comes to building a positioning system with accuracy in some centimetre range. A key challenge is how to cope with propagation delays internal to the wireless communication devices, sometime also called transmit and receive antenna delays. These delays must be calibrated out as a part of the synchronization scheme.
As shown in Fig. 3a, the delay on the transmit side denotes the time between the capturing
of the transmit time-stamp of the packet to the time when the first sample of the packet leaves
the transmit antenna. On the receive side, the delay indicates the time when the first sample
of the packet enters the receive antenna and the instance when the receive time-stamp is
captured in the based-band processor. In [
It is crucial to estimate and compensate the propagation delays internal to the clock master and the slaving device. In the AGL system, similar to the master, the slave node also transmits navigation packets on the UWB channel. In the th packet, the transmit time-stamp translated This computed value is feedback from node to in the next packet from . The node uses the classical proportional-integral (PI) control loop to track the this delay . Fig.3b shows this loop filter. The output of this filter is the amount of time by which the transmit time-stamp of the next packet from should be adjusted to compensate the internal delays.
Using the above elements for estimation and tracking of clock parameters and internal
propagation delays, we implemented a multi-hop synchronization scheme on a UWB test-bed using
radio transceiver from Qorvo, compliant with IEEE 802.15.4-2011 standard [
In the self-positioning mode, the tag can compute and track its position based on the pseudorange measurements with respect to multiple AGL anchors, Fig. 5. For this position computation, there is well established array of signal processing tools and methods including variants of Kalman filtering.
These pseudoranges are estimated by the tag from the receive timestamp of the navigation packets broadcasted by the anchors, and the corresponding transmit timestamps. Since the transmission of these packets occurs sequentially in a TDMA frame, the tag clock drifts during this period. To ensure accurate positioning of the tag, it is essential to compensate for this drift. The AGL tag incorporates clock drift compensation mechanisms, based on a quadratic clock model in (1), which enable it to accurately estimate its position.
The pseudorange from the tag to an anchor can be expressed as shown in equation (4): on the master ′ is used in combination with the ToF to the master clock is included, denoted as ′, see Fig. 3b. The corresponding receive time-stamp value (computed from the known positions of the two devices) to estimate the internal propagation delay. It can be shown that ≜ ′ − ′ − = , + , + ,
+ , . = ⏟√⏟⏟(⏟⏟⏟⏟−⏟⏟⏟⏟⏟)⏟2⏟⏟+⏟⏟⏟(⏟⏟⏟⏟−⏟⏟⏟⏟)⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟ ⏟⏟)2 + + .
2 + ( − ≜ Here, represents the pseudorange, denotes the actual range, and is the ofset between the anchor reference clock and the tag clock. The error in the range measurement, denoted by in (3) ′ (4) (a) Test setup comprising anchor nodes and a test node . Clock is distributed from node 0 to remaining nodes over the wireless channel. The node periodically sent blink messages. Sync error calculated based on receive time-stamp of these messages at anchor nodes. (4), is mainly caused by clock synchronization vagaries ( ), internal propagation delays ( ), and the impact of multipath and NLOS on the direct/first path detection ( ).
When computing the tag position, same error in the pseudoranges of the anchors does not have a significant impact on the accuracy. Most positioning algorithms can handle this type of error and it can be included as an additional unknown in the set of equations, lumped together with the ofset term
. However, it is crucial to minimize the variance in the error term across anchors. In the edge spaces, such as in a factory or onboard a ship, there could be a lot of metallic structures in close proximity causing heavy multipath and NLOS propagation. These wireless channel vagaries are the leading contributors to this variance. The accuracy and usefulness of a UWB positioning system depend on its ability to mitigate these impacts. To this end, the AGL tag implements a robust multi-layered multipath mitigation algorithm leveraging channel impulse response, receiver noise statistics and dynamics of pseudo ranges. The details of this algorithm will be published elsewhere in a separate publication. In the following sections, we present the performance evaluation of the AGL system in various test environments.
To measure the accuracy of the positioning system, we define two types of errors: horizontal and vertical. The 2D horizontal error for a position estimate (, ̂ )̂ is defined as the distance between the estimated location of the tag reported by the positioning system and the actual position of the tag, i.e., ℎ = √( ̂ − )2 + ( ̂ − )2, where ( , ) is the corresponding true position of the tag, also known as the ground truth. The cumulative distribution function (CDF) of the positioning error is defined as CDF() = ( ℎ < ). The vertical or height estimation error root-mean squre (RMS) value for the horizontal and vertical errors are defined as and corresponding CDF are defined as = | ̂ − | and CDF( ) = ( < ), respectively. The
The first set of tests were conducted in a sports-hall at Royal Military Academy Belgium. We evaluated the positioning accuracy in various system configurations. In this section, we present the results from one of the test configurations. Fig. 6a shows the test setup, where we deployed six anchors over an 18 m-by-9 m area. Among the six anchors, four were installed at a height of 3 m, while the remaining two were placed at a height of 0.25 m from the ground. We also tested the converse configuration and observed similar accuracy results.
To evaluate the accuracy of the AGL system, we set up a number of reference points within the anchor deployment area, as shown in Fig. 6b. At these test points, we logged the position estimates from the AGL tag and computed the empirical CDF for the 2D horizontal and vertical errors. highlights the RMS error value as well as the error at most commonly used reference points such as 50th, 68th and 95th percentile points.
Table 1 provides a summary of the horizontal error statistics for test points T0 to T10 in the sports hall setup. Notably, the worst-case error across all test points averages at 20 cm and the RMS error is approximately half of that. These results demonstrate the accuracy of the AGL ℎ, = √ 1 =1 ∑ 2 , , = ℎ √
1 =1 ∑ 2 . (5) (a) Test setup in the sports-hall.
(b) Tag estimated and true positions at a number of test points within the anchor deployment area. system and its consistency across multiple test points. The error variance that we see across the test points is mainly due to the geometric dilution of precision.
Fig. 7b shows the empirical CDF for the height error at test point T7. The summary statistics for the height error across all test points are presented in Table 2. On average, the worst-case error is limited to approximately 53 cm, the RMS error is slightly higher than half of this value, and the median error is approximately half of the maximum observed error. The vertical accuracy is contingent upon the deployment configuration and is not inherently constrained by the system itself. The accuracy in the vertical dimension can be further improved, for instance,by deploying more anchors and at three diferent heights instead of two.
We also conducted tests onboard a ship to access how well the AGL system is able to cope with the multipath in an environment with a lot of metallic structures in close proximity. We deployed six anchors on the ship deck, on easily reachable points, see Fig. 8. To gauge positioning accuracy, several test points were set up, and the average results across these reference points are tabulated in Table 3. Despite the challenging wireless conditions, we can 100 90 80 70 ) 60 % (F50 D C40 30 20 10 0
RMS:27.8cm
95% 68% 50% see that the AGL system demonstrated good accuracy, showcasing its resilience in such harsh environments.
The accuracy benchmarking with static test points has its limits because it does not capture and quantify the impact of dynamic scenarios where the tag is in motion. To address this gap, we conducted a benchmarking study using a visual motion capture system at one of the ID2Move test facilities in Nivelles, Belgium. The visual motion capture system is from Qualisys and consists of 12 cameras with 3 cameras installed on either side of a rectangular test zone. Fig. 9a shows two of these cameras in view. This Qualisys system can capture the exact position of the robotic devices with six degrees of freedom (6DOF) and in real-time with an accuracy better than a couple of centimeters. For this benchmarking study, we installed the AGL system comprising 6 anchors, shown in Fig. 9a. The tag was mounted on an unmanned ground vehicle (UGV). The UGV or rover is equipped with a Pixhawlk controller running Ardupilot software.
Fig. 9b illustrates the rover setup used in this benchmarking study. The setup included the AGL tag, several markers for visual tracking of the rover, and a Raspberry Pi for logging
a0 a5 .c732 RMS:15.4cm m 5 10 15 Horizontalerror(cm) 25 20 30
35 (b) Horizontal error CDF in R1 round.
a2 a3 a4 -600-600 -400 -200 0 x 2(c0m0) 400 600 800 1000 (a) Traversed path by the rover in R1 round from the AGL
system and the QTM motion capture system. data from the AGL tag and the Qualisys Track Manager (QTM). All the cameras are connected to the QTM server where the 6DOF position is calculated. The server provides an SDK for streaming the position data in real-time over Ethernet or WiFi. We used the Raspberry Pi’s WiFi connectivity to stream and log the position of the optical markers on the rover, as well as the position and orientation of the rover as a rigid body.
We conducted several rounds of tests with the rover, varying its speed and movement patterns. In the following sections, we present the results from three of these test rounds. Table 4 provides information on the average and maximum speed of the rover during these test runs.
To quantify the positioning accuracy of the AGL system, we used the position reported by the QTM server as the ground truth. We time-stamped, on the Raspberry Pi, the reception of data from both the AGL and QTM. For each AGL tag position data, we retrieved the corresponding data from the QTM log that was closest in time to the AGL time-stamp.
Fig. 10a shows the rover’s path during round R1, with positions reported by both the AGL tag and the QTM system, as well as the positions of the AGL anchors labeled 0 through 5. Overall the two traces closely follow each other with slightly increased divergence towards the edges of the AGL deployment zone. This divergence is mainly due to the geometric dilution of precision inherent in such a positioning system. To quantify the positioning error, we plot the horizontal error CDF in Fig. 10b. The RMS error is 15.4 cm and the maximum error observed during this round is limited to 32 cm. The error at 50th, 68th and 95th percentile points on the CDF curve is 14.3, 15.7 and 23.7 cm, respectively. While these errors are slightly higher than those observed in the static scenario in the previous section, the accuracy is still quite good.
95% 100 90 80 70 ) 60 % (F50 DC40 30 20 10 0 5 100 90 80 70 ) 60 % (F50 CD40 30 20 10 0
68% 50%
Test round R2 results are presented in Figs. 11a and 11b, which show the path taken by the rover and the corresponding positioning error CDF, respectively. There is a slight increase in positioning error compared to round R1, which is reflected in the CDF in Fig. 11b. The error has increased on all percentile points but remains below 30 cm for almost 99% of the time. This increase can be attributed to both the higher speed of the rover during this round and the diferent path trajectory it followed. These choices were intentional in order to capture relevant accuracy degradation patterns.
The results for the test round R3 are shown in Fig. 12a and Fig. 12b. As shown in the first ifgure, the paths reported by the AGL tag and QTM system closely align with each other. In the second figure, the error CDF curve shows that the error remained below 30 cm for more than 99% of the time, and the RMS error was limited to 14.2 cm. It is noteworthy that, despite the highest speed of the rover in this run compared to the other two, the observed error is the lowest. This can be attributed to the fact that the rover traversed path remains relatively within the inner deployment zone of the anchors.
For easy referencing, Table 5 summarizes the results from this kinematic scenario.
The work presented in this paper provides valuable insights into the accuracy and performance of the Agilica AGL positioning system. The test results show that the AGL system can provide decimeter level accuracy. In a static scenario, the position error at the 50th and 95th percentile points is around 10 cm and 15 cm, respectively. The error at these reference points respectively increased to about 15 cm and 25 cm, when the tag is moved with an average speed between 4 km/h to 7 km/h and maximum speed between 7 km/h and 11 km/h. This shows that the system can maintain high accuracy in kinematic scenarios where the tag is mobile.
The accuracy results have also been consistent across diferent test rounds, indicating the system reliability and robustness. The results also highlighted the impact of factors such as the reference point location, tag movement speed and trajectory on positioning accuracy, as well as the inherent limitations of the positioning system due to geometric dilution of precision. As such, it is important for users of such a positioning system to be aware of these factors and make appropriate choices in terms of deployment and usage to ensure optimal performance.
The paper has demonstrated the potential of the AGL system for accurate positioning in static as well as dynamic scenarios, while also highlighting the need for careful consideration of various factors that can impact accuracy, such as need for tight time synchronization, internal propagation delays compensation and the importance of minimizing the impact of multipath propagation. These results can be useful for researchers and practitioners in fields such as robotics, autonomous vehicles, and indoor positioning, where accurate and reliable positioning is essential for a range of applications.
The AGL system is built as an alternative and complementary to the GNSS for applications in the edge spaces that requires high accuracy and availability. However, it should be highlighted that when it comes to the alternatives to the GNSS, a single solution, technology or system cannot cater to the needs in all operational environments. It would be more feasible and sensible to create customized independent systems that are designed to function optimally within a particular environment or group of environments. However, these systems should have a standardized and compatible application interface so that data from multiple systems can be merged together when available to form a unified positioning system. The integrated system would be more robust, precise, and dependable. This approach underpins the AGL system.