A Tutorial on Infrastructure Self-Calibration for Radio-Based Indoor Positioning

arXiv:2611.15503 · eess.SP · Submitted 2026-11-11 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "A Tutorial on Infrastructure Self-Calibration for Radio-Based Indoor Positioning".

Dev: Here's the gist of the paper. Radio positioning requires knowing the locations of infrastructure elements such as anchors, access points, base stations, or reconfigurable intelligent surfaces; manual surveying is costly and error-prone.

Rosa: Calibrating this infrastructure using its own measurements is framed as an estimation problem.

Dev: The tutorial addresses this problem by presenting a single factor graph approach that incorporates measurement topology into its factor structure, reference information as priors, and other measurement models as nuisance states.

Rosa: It demonstrates what different ranges of information can identify, from system rigidity to the required motion contribution from a mobile device.

Title and authors: Rosa: So we're looking at this paper called "A Tutorial on Infrastructure Self-Calibration for Radio-Based Indoor Positioning". It seems like it tackles how you can figure out where all your radio stuff is—like anchors or base stations—just by using measurements they are already making, instead of having people come out and do costly surveys.

Dev: Right. Basically, the core idea is turning that tricky positioning problem into an estimation problem where the infrastructure calibrates itself using its own data. It frames this whole process as a single factor graph that accounts for how those measurements are laid out in space, which is pretty clever because it ties the measurement topology directly into the structure of the math.

Taro: I'm curious about what kind of information this setup can actually pull out from just these measurements. It sounds like they’re trying to figure out if they can get things as rigid as possible or maybe even estimate how much motion a mobile device has to contribute to the fix.

Rosa: Exactly, and it shows you what different amounts of information you have—like the reference information you start with—can identify, ranging from just how rigid the whole system is to figuring out the required movement from a mobile device. They use things like multidimensional scaling, nonlinear least squares, and recursive filters as their engines for this estimation problem.

Dev: And what I find interesting is how they structure it: they treat the measurement topology as defining the factors of this graph, and then reference information acts as priors to give the system a starting point. Other measurement models get treated like nuisance states so they don't spoil the main goal of finding those positions.

Taro: That sounds like a solid way to model it, but I wonder how robust this is when things get messy in the real world, for instance with non-line-of-sight propagation or when you have partial networks. They test this on two different setups: one with a full mesh and another that has some non-line-of-sight issues and a public partial mesh.

Rosa: They do, and those tests lead to three key lessons for practice regarding how anchor errors are affected depending on the network structure, and they also discuss range offsets in line of sight deployments. It’s practical stuff they pull out from running those examples with measured data published separately as a dataset.

Dev: So, the paper isn't just theoretical; it’s grounded in actual measurements from two different real-world scenarios to show what works and what doesn't when you have different network layouts. That gives the reader something concrete to work with instead of just abstract concepts about factor graphs.

Taro: I mean, if we think about the world for someone who only listens to this, it means they can move past just needing perfect knowledge of where everything is beforehand and start building a system that figures out its own geometry as it operates in the field.

Rosa: That’s right. The tutorial guides you from knowing nothing to deciding on a specific topology, picking the right estimation engine for your needs, and figuring out exactly what kind of calibration checks you actually need to do. It’s a decision path from deployment to method that’s pretty useful for anyone trying to set up a system.

Dev: And they present this guidance through a decision tree structure, which prompts questions about things like whether the mobile device is also ranging and if the thinned graph is globally rigid, helping you choose the right approach. It makes it less of a black box and more of a systematic process.

Taro: I wonder how useful that decision tree really is when things go completely sideways, like when you have severe noise or bias in your measurements; does it still give you a good recovery path?

Rosa: Well, the paper touches on some ways to make this system more robust, and one improvement they suggest is letting the loss function—how the system decides what kind of error to prioritize—be chosen dynamically based on deployment characteristics to handle things like non-line-of-sight biases or range outliers better.

Dev: That’s a big deal because when you have noisy or obstructed measurements, a fixed kernel might just fail, but adapting the kernel based on what the data looks like seems like it could help maintain accurate position estimates. It’s about tuning the filter itself to handle the noise distribution you observe.

Taro: If we look at mobile-only topologies where a device is moving through space, they seem to have ways to optimize those motion factors so that the system doesn't take too many iterations or epochs to actually converge on a solution. That directly impacts how fast you can get a fix while the device is walking.

Rosa: They do, by showing that incorporating motion factors into those algorithms can reduce the number of epochs required for convergence, which means better performance during tag walks. It’s about making the estimation process faster in dynamic scenarios.

Dev: Another improvement they point to is using an external frame, like a map alignment or a known reference system, to fix the absolute frame when you don't have enough initial seeds or reliable poses. They say that an external frame needs a certain number of seeds or referenced poses to anchor the estimate correctly.

Taro: That makes sense for getting absolute grounding, but I wonder if that external alignment still works well when the local environment itself is changing rapidly, like during a fast movement sequence.

Rosa: The paper shows that this external frame approach helps when seeds are insufficient or unreliable, but it also points out an open problem: how much systematic bias and adaptive noise scales really remain after you've done the offset calibration. It suggests tuning kernels based on the residual distribution to address those biases without needing ground truth data.

Dev: That’s a fair critique; you can calibrate the local geometry, but you still have that unknown systematic error lurking in the measurement models themselves, and adaptive noise scales could be a way to tackle that uncertainty. It moves away from using fixed kernels for every scenario.

Taro: And on top of all this calibration stuff, they suggest an idea for checking integrity: running the estimation on smaller, redundant subgraphs of the sensor network to detect edge-level faults in the system.

Rosa: So, to wrap up, this tutorial on infrastructure self-calibration shows that by structuring it as a single factor graph and using reference information as priors, you can cover everything from just checking system rigidity to estimating motion contributions. The examples with UWB two-way ranging in 2D contrast full meshes with partial meshes to give you practical lessons <ref:2611.15503#pg2>.

Dev: Ultimately, the paper gives practitioners a path: they show you how to define the topology, select the inference engine based on your needs and the topology, and determine what calibration checks are necessary for your specific setup. It’s a systematic way to handle radio positioning problems without needing constant manual surveying.

Taro: For someone listening who only cares about real-world application, it means you can design systems that don't just rely on perfect initial conditions but actively use the measurements they collect to refine their own understanding of the environment.

Rosa: That’s the gist of it. The tutorial on infrastructure self-calibration for radio-based indoor positioning is a deep dive into turning complex physical placement problems into structured estimation problems solved by factor graphs. It gives you a framework for building robust systems that can calibrate themselves using their own data across all those different measurement topologies.

The paper's summary: Rosa: So, this paper is basically showing how you can figure out where all your radio stuff is—like anchors or base stations—just by using measurements they are already making instead of doing costly surveys to find them manually.

Dev: Right, it frames the whole location problem as an estimation problem where the infrastructure calibrates itself using its own data. It uses a single factor graph that ties the measurement layout directly into the structure of the math, which is pretty smart because it links how things are measured in space to how we model them.

Taro: I'm curious about what kind of information this setup can actually pull out from just these measurements. Does it help you get something as rigid as possible, or maybe estimate how much movement a mobile device has to contribute to the fix?

Rosa: It shows you what different amounts of information you have—like the starting reference info you begin with—can identify, ranging from just how rigid the whole system is to figuring out the required motion contribution from a mobile device. They use things like multidimensional scaling, nonlinear least squares, and recursive filters as their engines for this estimation problem.

Dev: And what I find interesting is how they structure it: they treat the measurement layout as defining the factors of this graph, and then reference information acts as priors to give the system a starting point. Other measurement models get treated like nuisance states so they don't spoil the main goal of finding those positions.

Taro: That sounds like a solid way to model it, but I wonder how robust this is when things get messy in the real world, for instance with non-line-of-sight propagation or when you have partial networks. They test this on two different setups: one with a full mesh and another that has some non-line-of-sight issues and a public partial mesh.

Rosa: They do, and those tests lead to three key lessons for practice regarding how anchor errors are affected depending on the network structure, and they also discuss range offsets in line of sight deployments. It’s practical stuff they pull out from running those examples with measured data published separately as a dataset.

Dev: So, the paper isn't just theoretical; it’s grounded in actual measurements from two different real-world scenarios to show what works and what doesn't when you have different network layouts. That gives the reader something concrete to work with instead of just abstract concepts about factor graphs.

Taro: I mean, if we think about the world for someone who only listens to this, it means they can move past just needing perfect knowledge of where everything is beforehand and start building a system that figures out its own geometry as it operates in the field.

Rosa: That’s right. The tutorial guides you from knowing nothing to deciding on a specific topology, picking the right estimation engine for your needs, and figuring out exactly what kind of calibration checks you actually need to do. It’s a decision path from deployment to method that’s pretty useful for anyone trying to set up a system.

Dev: And they present this guidance through a decision tree structure, which prompts questions about things like whether the mobile device is also ranging and if the thinned graph is globally rigid, helping you choose the right approach. It makes it less of a black box and more of a systematic process.

Taro: I wonder how useful that decision tree really is when things go completely sideways, like when you have severe noise or bias in your measurements; does it still give you a good recovery path?

Rosa: Well, the paper touches on some ways to make this system more robust, and one improvement they suggest is letting the loss function—how the system decides what kind of error to prioritize—be chosen dynamically based on deployment characteristics to handle things like non-line-of-sight biases or range outliers better.

Dev: That’s a big deal because when you have noisy or obstructed measurements, a fixed kernel might just fail, but adapting the kernel based on what the data looks like seems like it could help maintain accurate position estimates. It’s about tuning the filter itself to handle the noise distribution you observe.

Taro: If we look at mobile-only topologies where a device is moving through space, they seem to have ways to optimize those motion factors so that the system doesn't take too many iterations or epochs to actually converge on a solution. That directly impacts how fast you can get a fix while the device is walking.

Rosa: They do, by showing that incorporating motion factors into those algorithms can reduce the number of epochs required for convergence, which means better performance during tag walks. It’s about making the estimation process faster in dynamic scenarios.

Dev: Another improvement they point to is using an external frame, like a map alignment or a known reference system, to fix the absolute frame when you don't have enough initial seeds or reliable poses. They say that an external frame needs a certain number of seeds or referenced poses to anchor the estimate correctly.

Taro: That makes sense for getting absolute grounding, but I wonder if that external alignment still works well when the local environment itself is changing rapidly, like during a fast movement sequence.

Rosa: The paper shows that this external frame approach helps when seeds are insufficient or unreliable, but it also points out an open problem: how much systematic bias and adaptive noise scales really remain after you've done the offset calibration. It suggests tuning kernels based on the residual distribution to address those biases without needing ground truth data.

Dev: That’s a fair critique; you can calibrate the local geometry, but you still have that unknown systematic error lurking in the measurement models themselves, and adaptive noise scales could be a way to tackle that uncertainty. It moves away from using fixed kernels for every scenario.

Taro: And on top of all this calibration stuff, they suggest an idea for checking integrity: running the estimation on smaller, redundant subgraphs of the sensor network to detect edge-level faults in the system.

Rosa: So, to wrap up, this tutorial on infrastructure self-calibration shows that by structuring it as a single factor graph and using reference information as priors, you can cover everything from just checking system rigidity to estimating motion contributions. The examples with UWB two-way ranging in 2D contrast full meshes with partial meshes to give you practical lessons <ref:2611.15503#pg2>.

Dev: Ultimately, the paper gives practitioners a path: they show you how to define the topology, select the inference engine based on your needs and the topology, and determine what calibration checks are necessary for your specific setup. It’s a systematic way to handle radio positioning problems without needing constant manual surveying.

Taro: For someone listening who only cares about real-world application, it means you can design systems that don't just rely on perfect initial conditions but actively use the measurements they collect to refine their own understanding of the environment.

Rosa: That’s the gist of it. The tutorial on infrastructure self-calibration for radio-based indoor positioning is a deep dive into turning complex physical placement problems into structured estimation problems solved by factor graphs. It gives you a framework for building robust systems that can calibrate themselves using their own data across all those different measurement topologies.

Dev: That framework really shifts the focus from just getting one perfect location to building a system that manages uncertainty and self-correcting over time in real-time loops. And next time, we’ll look at how they handle those specific noise distributions when things get biased.

The paper's improvements: Rosa: So, this paper is showing how to fix the problem of knowing where radio infrastructure is by letting the infrastructure itself do the calibration using its own measurements instead of having people come out and do surveys.

Dev: It uses a single factor graph that accounts for how those measurements are laid out in space, which links the measurement topology directly into the structure of the math. This is a smart way to model it because it ties how things are measured in space to how we model them.

Taro: I'm curious about what kind of information this setup can actually pull out from just these measurements. Does it help you get something as rigid as possible, or maybe estimate how much movement a mobile device has to contribute to the fix?

Rosa: It shows you what different amounts of information you have—like the starting reference info you begin with—can identify, ranging from just how rigid the whole system is to figuring out the required motion contribution from a mobile device. They use things like multidimensional scaling, nonlinear least squares, and recursive filters as their engines for this estimation problem.

Dev: And what I find interesting is how they structure it: they treat the measurement layout as defining the factors of this graph, and then reference information acts as priors to give the system a starting point. Other measurement models get treated like nuisance states so they don't spoil the main goal of finding those positions.

Taro: That sounds like a solid way to model it, but I wonder how robust this is when things get messy in the real world, for instance with non-line-of-sight propagation or when you have partial networks. They test this on two different setups: one with a full mesh and another that has some non-line-of-sight issues and a public partial mesh.

Rosa: They do, and those tests lead to three key lessons for practice regarding how anchor errors are affected depending on the network structure, and they also discuss range offsets in line of sight deployments. It’s practical stuff they pull out from running those examples with measured data published separately as a dataset.

Dev: So, the paper isn't just theoretical; it’s grounded in actual measurements from two different real-world scenarios to show what works and what doesn't when you have different network layouts. That gives the reader something concrete to work with instead of just abstract concepts about factor graphs.

Taro: I mean, if we think about the world for someone who only listens to this, it means they can move past just needing perfect knowledge of where everything is beforehand and start building a system that figures out its own geometry as it operates in the field.

Rosa: That’s right. The tutorial guides you from knowing nothing to deciding on a specific topology, picking the right estimation engine for your needs, and figuring out exactly what kind of calibration checks you actually need to do. It’s a decision path from deployment to method that’s pretty useful for anyone trying to set up a system.

Dev: And they present this guidance through a decision tree structure, which prompts questions about things like whether the mobile device is also ranging and if the thinned graph is globally rigid, helping you choose the right approach. It makes it less of a black box and more of a systematic process.

Taro: I wonder how useful that decision tree really is when things go completely sideways, like when you have severe noise or bias in your measurements; does it still give you a good recovery path?

Rosa: Well, the paper touches on some ways to make this system more robust, and one improvement they suggest is letting the loss function—how the system decides what kind of error to prioritize—be chosen dynamically based on deployment characteristics to handle things like non-line-of-sight biases or range outliers better.

Dev: That’s a big deal because when you have noisy or obstructed measurements, a fixed kernel might just fail, but adapting the kernel based on what the data looks like seems like it could help maintain accurate position estimates. It’s about tuning the filter itself to handle the noise distribution you observe.

Taro: If we look at mobile-only topologies where a device is moving through space, they seem to have ways to optimize those motion factors so that the system doesn't take too many iterations or epochs to actually converge on a solution. That directly impacts how fast you can get a fix while the device is walking.

Rosa: They do, by showing that incorporating motion factors into those algorithms can reduce the number of epochs required for convergence, which means better performance during tag walks. It’s about making the estimation process faster in dynamic scenarios.

Dev: Another improvement they point to is using an external frame, like a map alignment or a known reference system, to fix the absolute frame when you don't have enough initial seeds or reliable poses. They say that an external frame needs a certain number of seeds or referenced poses to anchor the estimate correctly.

Taro: That makes sense for getting absolute grounding, but I wonder if that external alignment still works well when the local environment itself is changing rapidly, like during a fast movement sequence.

Rosa: The paper shows that this external frame approach helps when seeds are insufficient or unreliable, but it also points out an open problem: how much systematic bias and adaptive noise scales really remain after you've done the offset calibration. It suggests tuning kernels based on the residual distribution to address those biases without needing ground truth data.

Dev: That’s a fair critique; you can calibrate the local geometry, but you still have that unknown systematic error lurking in the measurement models themselves, and adaptive noise scales could be a way to tackle that uncertainty. It moves away from using fixed kernels for every scenario.

Taro: And on top of all this calibration stuff, they suggest an idea for checking integrity: running the estimation on smaller, redundant subgraphs of the sensor network to detect edge-level faults in the system.

Rosa: So, to wrap up, this tutorial on infrastructure self-calibration shows that by structuring it as a single factor graph and using reference information as priors, you can cover everything from just checking system rigidity to estimating motion contributions. The examples with UWB two-way ranging in 2D contrast full meshes with partial meshes to give you practical lessons <ref:2611.15503#pg2>.

Dev: Ultimately, the paper gives practitioners a path: they show you how to define the topology, select the inference engine based on your needs and the topology, and determine what calibration checks are necessary for your specific setup. It’s a systematic way to handle radio positioning problems without needing constant manual surveying.

Taro: For someone listening who only cares about real-world application, it means you can design systems that don't just rely on perfect initial conditions but actively use the measurements they collect to refine their own understanding of the environment.

Rosa: That’s the gist of it. The tutorial on infrastructure self-calibration for radio-based indoor positioning is a deep dive into turning complex physical placement problems into structured estimation problems solved by factor graphs. It gives you a framework for building robust systems that can calibrate themselves using their own data across all those different measurement topologies.

Dev: That framework really shifts the focus from just getting one perfect location to building a system that manages uncertainty and self-correcting over time in real-time loops. And next time, we’ll look at how they handle those specific noise distributions when things get biased.

Conclusion: Rosa: So, to wrap up, this tutorial on infrastructure self-calibration for radio-based indoor positioning shows how you can figure out where all your radio stuff is by letting the infrastructure itself do the calibration using its own measurements instead of having people come out and do surveys.

Dev: It uses a single factor graph that accounts for how those measurements are laid out in space, which links the measurement topology directly into the structure of the math. This is a smart way to model it because it ties how things are measured in space to how we model them.

Taro: For someone listening who only cares about real-world application, it means you can design systems that don't just rely on perfect initial conditions but actively use the measurements they collect to refine their own understanding of the environment.

Rosa: That’s right. The tutorial on infrastructure self-calibration for radio-based indoor positioning is a deep dive into turning complex physical placement problems into structured estimation problems solved by factor graphs. It gives you a framework for building robust systems that can calibrate themselves using their own data across all those different measurement topologies.

Dev: That framework really shifts the focus from just getting one perfect location to building a system that manages uncertainty and self-correcting over time in real-time loops. And next time, we’ll look at how they handle those specific noise distributions when things get biased.

Taro: I just think the way they tackle those potential failures by checking redundant subgraphs is really important for autonomy research because it builds an integrity layer into the system itself.

Rosa: It does, and that idea of checking for edge-level faults in smaller, redundant sensor networks is a practical safety net that you can actually use when deploying these systems in the field.

Dev: Yeah, so if you’re building this for real-time deployment, having a clear path from deployment to method defined by that decision tree is key to making sure the system stays within its safety case.

Taro: It makes sense because you don't want an AI trying to figure out which measurement primitive or algorithm family to use on the fly without a plan in place.

Rosa: Exactly. So, this paper gives us a solid method for turning physical placement problems into estimation problems solved by factor graphs, but there’s still some work left on those systematic biases and adaptive noise scales we discussed earlier.

Dev: Right, that leaves a lot of room to explore how we can make these solutions even more resilient when the world is truly unpredictable or noisy.

TUD Dresden University of Technology

eess.SP

Submitted: 2026-11-11

Updated: 2026-11-11

Code: https://github.com/PaulSchwarzbach/radiocal

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: Radio positioning requires knowing the locations of infrastructure elements such as anchors, access points, base stations, or reconfigurable intelligent surfaces…

Key concepts

Factor Graph Approach
This is a mathematical framework used to model complex estimation problems, like radio calibration. It treats all measurements and constraints as 'factors' in a graph structure. This allows the system to simultaneously solve for unknown variables (like infrastructure positions) by balancing all available data points.
Measurement Topology
This refers to the physical arrangement of radio elements in an environment, such as whether they form a fully connected mesh or have non-line-of-sight links. The paper shows how the chosen mathematical model must change depending on this topology to accurately estimate positions.
Reference Information (Priors)
These are pieces of prior knowledge used in the estimation process, such as known locations or constraints about device motion. They help guide the system when measurements are incomplete, allowing it to recover geometry even with limited data.

Terminology

Summary

Radio positioning requires knowing the locations of infrastructure elements such as anchors, access points, base stations, or reconfigurable intelligent surfaces; manual surveying is costly and error-prone. p. 0

Where the paper says it

1 A Tutorial on Infrastructure Self-Calibration for Radio-Based Indoor Positioning Paul Schwarzbach Abstract—Radio positioning needs the positions of its anchors, access points, base stations, or reconfigurable intelligent surfaces, and manual surveying is costly and error-prone. Calibrating this infrastructure from its own measurements is one estimation problem, studied in at least four communities as self-calibration, auto-positioning, anchor self-localization, network localization, or radio simultaneous localization and mapping.

Calibrating this infrastructure using its own measurements is framed as an estimation problem. p. 0
Where the paper says it

1 A Tutorial on Infrastructure Self-Calibration for Radio-Based Indoor Positioning Paul Schwarzbach Abstract—Radio positioning needs the positions of its anchors, access points, base stations, or reconfigurable intelligent surfaces, and manual surveying is costly and error-prone. Calibrating this infrastructure from its own measurements is one estimation problem, studied in at least four communities as self-calibration, auto-positioning, anchor self-localization, network localization, or radio simultaneous localization and mapping.

The tutorial addresses this problem by presenting a single factor graph approach that incorporates measurement topology into its factor structure, reference information as priors, and other measurement models as nuisance states. p. 0
Where the paper says it

This tutorial teaches the problem through one factor graph. The measurement topology sets its factors, the reference information enters as priors, and other measurement models add nuisance states.

It demonstrates what different ranges of information can identify, from system rigidity to the required motion contribution from a mobile device. p. 0
Where the paper says it

The measurement topology sets its factors, the reference information enters as priors, and other measurement models add nuisance states. The tutorial shows what ranges can identify, from rigidity to the motion a mobile device must contribute, and presents multidimensional scaling, nonlinear least squares, recursive filters, and convex relaxations as inference engines over this graph.

The inference engines employed include multidimensional scaling, nonlinear least squares, recursive filters, and convex relaxations. p. 0
Where the paper says it

The measurement topology sets its factors, the reference information enters as priors, and other measurement models add nuisance states. The tutorial shows what ranges can identify, from rigidity to the motion a mobile device must contribute, and presents multidimensional scaling, nonlinear least squares, recursive filters, and convex relaxations as inference engines over this graph.

The paper makes three primary contributions:

  1. A unified formulation across different measurement topologies: Range-based calibration is treated as one factor graph where the factor structure follows the measurement topology and priors carry reference information, while other primitives introduce nuisance states. p. 1

    Where the paper says it

    • One formulation across communities. Range-based calibration is stated as one factor graph, whose factor structure follows the measurement topology and whose priors carry the reference information, while other primitives add nuisance states.

  2. A decision path from deployment to method: A decision tree guides a practitioner from initial knowledge to defining the topology, selecting an inference engine, and determining necessary calibration checks. p. 1

    Where the paper says it

    • A path from deployment to method. A decision tree leads from what a practitioner knows to a topology, an engine, and the checks a calibration needs, and each leaf names the examples it rests on (Section VII).

  3. Reference implementations on open data: The tutorial includes practical examples using UWB two-way ranging in two dimensions, contrasting a full mesh deployment with one featuring non-line-of-sight (NLOS) propagation and a public partial mesh. p. 2

    Where the paper says it

    3 • Reference implementations and examples on open measured data. One session of a 10-anchor UWB deployment is carried through every step, and a second deployment rich in non-line-of-sight (NLOS) propagation and a public partial mesh serve as contrasts.

The tutorial is structured into seven sections: Section II introduces the problem and running example; Section III formulates the problem as a factor graph; Section IV details what ranges can be identified by the graph; Section V presents the inference engines; Section VI runs these engines on measured data; and finally, Section VII converts the results into a decision tree. p. 2

Where the paper says it

Section III formulates the problem as one factor graph, Section IV shows what its ranges can identify, and Section V presents the inference engines. Section VI runs them on measured data, and Section VII turns the results into a decision tree.

The tutorial discusses various measurement topologies: Fully meshed, Multi-hop, Mobile-assisted, and Mobile-only. It also classifies calibration regimes based on reference information: Blind calibration (K = K m =), Semi-blind calibration (1 at most K < N), and Non-blind calibration (K = N). The examples utilize UWB two-way ranging in a two-dimensional space. p. 3

Where the paper says it

Arrival times, angles, and multipath components are extracted from the channel fully meshed multi-hop mobileassisted mobileonly measurement topology blind K = Km = ∅ mobilereferenced K = ∅ ̸= Km semi-blind 1 ≤ |K| < N non-blind |K| = N reviews and tutorials reference information UWB WiFi BLE cellular acoustic signals of opportunity other radio several technologies or agnostic no mobile node no anchor position left to estimate [16] [27] [22] [39] [74] [75] [70] [71] [36] [69] [11] [76] [42] [77] [14] [26] [78] [79] [37] [68] [61] [9] [19] [41] [80] [13] [3] [58] [55] [49] [48] Fig. The two-dimensional classification space with works for further study.

Three key lessons are drawn from contrasting full mesh versus thinned/partial meshes, two lessons address range offsets in line-of-sight deployments, and deployment guidance is derived from the decision tree. p. 2
Where the paper says it

One session of a 10-anchor UWB deployment is carried through every step, and a second deployment rich in non-line-of-sight (NLOS) propagation and a public partial mesh serve as contrasts. Section VI draws three lessons for practice from them, each with its effect on the anchor error.

The decision tree prompts questions regarding deployment characteristics: whether a mobile device also ranges to anchors, if the device moves through space, if all anchor pairs range directly, and whether the thinned graph is globally rigid. p. 2

Where the paper says it

Mobile-assisted. A device ranges to the infrastructure on top of the inter-anchor links, whether it moves along a path or stands at a few points, so the edge set is mixed, E ⊆ (Va × Vm) ∪ Eaa with Eaa ̸= ∅ [64], [68].

The examples test neither how anchor errors affect position fixes nor offset calibration on surveyed networks; instead, accuracy targets set the bar for anchor errors, and a survey leaves range offsets of primitives (Table IV) as nuisance states.

Open problems identified include Systematic Bias and Adaptive Noise Scales, Testing Identifiability in Sparse and Dynamic Views, Beyond Two-Dimensional Two-Way Ranging, and Learning-Based Methods and Their Evaluation. The code, result tables, data, and literature records are openly released under a CC BY 4.0 license. The tutorial concludes by summarizing that calibrating radio infrastructure from its own measurements is fundamentally one estimation problem treated as a single factor graph across various measurement topologies and reference information types. p. 0

Where the paper says it

Calibrating this infrastructure from its own measurements is one estimation problem, studied in at least four communities as self-calibration, auto-positioning, anchor self-localization, network localization, or radio simultaneous localization and mapping. This tutorial teaches the problem through one factor graph.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Robust loss selection can be dynamically chosen based on deployment characteristics to handle NLOS propagation biases and range outliers effectively. This allows AI systems to maintain accurate position estimates even when measurements are noisy or obstructed, as seen in Section VI-A where the cauchy kernel removes most of a shift under NLOS distortion.

  2. Deployment guidance can be derived from a decision tree that selects the appropriate measurement primitive, node roles, and algorithm family based on the deployment topology. This provides a systematic path from what a practitioner knows to a topology, an engine, and the checks a calibration needs (Section VII).

  3. AI systems can perform blind calibration by using only range measurements without reference positions or trajectory priors. This capability allows for the recovery of geometry up to rigid transformation, as noted in Section IV-A where blind calibration (K = Km = ∅) leaves the full Euclidean invariance.

  4. The system can implement adaptive noise scales that tune kernels based on residual distribution, replacing fixed kernels. This adaptation is crucial for handling biased ranges without ground truth, addressing the open problem of how much bias remains after offset calibration (Section VIII-A).

  5. Motion factors in mobile-only topologies can be optimized to reduce the required number of epochs for convergence. This directly improves performance on tag walks by using motion factors reduce the epochs required (Section V).

  6. The system can utilize an external frame, such as a map alignment, to fix the absolute frame when seeds are insufficient or unreliable. This is achieved because an external frame needs d + 1 seeds, referenced poses, or a map alignment (Section VI-E) to anchor an estimate.

  7. The system can perform edge-level fault detection by running it on redundantly rigid subgraphs of sensor networks. This adds an integrity layer by testing the misspecification of Section VIII-F as one candidate (Section VIII).

Abstract

Radio positioning needs the positions of its anchors, access points, base stations, or reconfigurable intelligent surfaces, and manual surveying is costly and error-prone. Calibrating this infrastructure from its own measurements is one estimation problem, studied in at least four communities as self-calibration, auto-positioning, anchor self-localization, network localization, or radio simultaneous localization and mapping. This tutorial teaches the problem through one factor graph. The measurement topology sets its factors, the reference information enters as priors, and other measurement models add nuisance states. The tutorial shows what ranges can identify, from rigidity to the motion a mobile device must contribute, and presents multidimensional scaling, nonlinear least squares, recursive filters, and convex relaxations as inference engines over this graph. One session of a ten-anchor ultra-wideband deployment is carried through every step, and recordings from a second, harsher deployment and a public partial mesh serve as contrasts. The examples illustrate three lessons. On a full mesh, choose a robust loss and set its scale first, and calibrate with the engine of the positioning system, since the engines then reach similar errors. On thinned or partial meshes, test the result at surveyed check points, because the start decides which minimum a solver reaches. On a line-of-sight deployment, estimate range offsets, which lower the error floor that a systematic ranging bias sets and a robust loss cannot remove. A decision tree and a table of their effects turn them into deployment guidance. Code and result tables are openly released, and the data are published separately.

Sources

Related papers