Most robot mower programmes start the same way. A GNSS module, an IMU and wheel encoders feed a filter, and on an open test lawn the machine holds its stripes to within a couple of centimetres. Someone then asks whether a camera-based fusion sensor is worth paying for. For that lawn, probably not.
The question is whether the product will only ever see that lawn. Customers' properties have trees, hedges, sheds and neighbours' two-storey extensions, and each of them degrades satellite positioning in its own way. This post looks at the build-versus-buy decision for visual-inertial odometry (VIO), the technique of tracking motion from a camera and an IMU between satellite fixes. Mowers make a good case study. Unit costs are tight, volumes are high, and a large share of the work area sits somewhere inconvenient.
Where the In-House Stack Works
The case for building deserves a fair hearing. An RTK GNSS, IMU and wheel-odometry stack handles a large share of properties. Consumer comparisons of mower navigation put RTK in the large, open-sky category and note that it degrades under canopy and near structures [1]. For a sports field or a business park with a clear horizon, a home-grown RTK/INS filter is a sensible build. The inertial navigation techniques involved are well documented, and a competent team will have something working quickly.
If the work areas never need vision, there is no VIO decision to make. The rest of this post assumes they sometimes do.
They usually do. Along a hedgerow or under a mature oak, the RTK solution drops to float or disappears, and wheel odometry takes over. Wheel odometry drifts, and it drifts faster on wet grass. Competitors have noticed. Consumer flagships increasingly pair RTK with visual SLAM or VIO, and EGO describes its AURA-R2 as combining RTK, visual SLAM and visual-inertial odometry [2]. Whether to add vision is mostly settled. The open question is whose vision stack ends up in the product.
What Buying Gets You
Fixposition's PBX-A1 is a useful reference because the vendor targets this application directly and it is the company's current hardware. It is an IP69k-rated fusion unit, so it survives the pressure washer that every mower eventually meets. Its xFusion software combines GNSS and vision algorithms with an IMU, so the camera carries the track when satellites become unreliable [3].
Two details matter for mowers. The first is camera placement. Fixposition lists support for external stereo and mono cameras, mounted wherever they have a useful view [3]. A camera 15 cm off the ground on a residential robot and one on a commercial zero-turn have different jobs, and a sensor that fixes the camera position makes that harder. The second is integrity. In an interview, Fixposition engineers said the system reports confidence levels along with position and orientation, which they described as essential for certifiable autonomy [4]. A mower that knows when it is unsure can slow down or stop. One that does not know will keep mowing.
Vendor claims deserve the same scepticism as anyone else's. Fixposition describes the PBX-A1 as delivering its predecessor's capabilities with "elevated performance" [3], which is not a number. The numbers sit in the datasheet, which Fixposition publishes alongside the integration manual [5]. Read it against your own dead-reckoning error budget before reading anything else. Corrections stay the integrator's decision under either approach, and the trade-offs between network and PPP-style services are worth settling early.
There is more than one way to buy:
- The PBX-A1 is the complete unit for platforms that want a finished, sealed sensor.
- The xFusion A1 Module is a system-on-module for integration onto a custom board.
- xFusion Software is a licence to run the fusion engine on your own SoC, which sits on the boundary between build and buy.
The Opportunity Cost of the Engineers
Most of the cost of building VIO is not the VIO. Getting an open-source estimator to track a public dataset takes days. The real cost is what those engineers do not build in the meantime: coverage planning, blade and deck control, docking, obstacle handling, the app, fleet tooling and safety work. Those are the features a customer compares. Estimator quality only becomes visible when it fails.
There is a fair counterargument. If positioning is the product, or the basis of the company's valuation, then buying it means outsourcing the differentiator. A company selling an autonomy stack should own its state estimator, or understand it well enough to change it. For that company the opportunity cost runs the other way.
The practical exercise is to price the team, not the sensor. A VIO effort with two or three dedicated engineers for two years costs more than a lot of sensors, before counting what those engineers were not working on. Set that total against unit cost multiplied by realistic volume. At very high consumer volumes the arithmetic can eventually favour building. At commercial-mower volumes it often does not.
Scope Creep Has a Direction
Build projects in this area rarely fail because the first scope was wrong. They fail by accretion, and the sequence is predictable enough to plan around.
The project starts as GNSS, IMU and wheels. Canopy tests expose the gap, so a camera is added. The camera brings exposure control, camera-to-IMU timing and calibration. Then comes lens contamination, because grass clippings and dew degrade a calibration that was perfect on the bench. Visual tracking also needs to start cleanly from a standstill, which feature-based VIO handles poorly when there is little motion to work with [6].
The fusion layer then has to decide when each source is lying. A plausible float solution, a multipath-corrupted fix beside a garage wall and a visual track on three features all need different treatment. That leads into integrity and residual monitoring, which was not in the original plan. After that come the compute budget on a cost-reduced SoC, relocalisation after a week on the dock, and map consistency across seasons.
Each item is reasonable, and each is something a good engineer will want to fix. That is why it is scope creep and not a planning error. A vendor has already absorbed many of these steps. Whether it did so well for your platform is what evaluation should establish, but the schedule risk is no longer yours.
Field Hours Cannot Be Skipped
This is the part that does not compress. VIO looks good on curated datasets and worse outdoors. The research literature says so plainly: monocular and stereo setups lack robustness under challenging lighting, texture and motion blur [7], and feature-based estimators struggle with low-texture scenes and abrupt illumination changes [6]. A lawn offers all of these. The texture is uniform, the camera sits low and sees mostly grass, shadows move with the sun, and morning light can overexpose half the frame.
Robustness therefore comes from field hours more than from algorithms. Failure modes appear across seasons, soil conditions, grass species and sites the team did not choose. The long tail then sets the product's reputation. At one session per mower per day and a 99.9% success rate, a fleet of ten thousand produces roughly ten failed sessions every day. Some of those mowers will be stranded somewhere.
Even established players keep paying for this. Greenzie's 5.0 release in March 2026 centred on better handling of GPS variability, informed by field data from its commercial deployments [8].
A vendor with a diverse installed base accumulates field hours faster than any single-platform programme, and that has real value. The catch is that those hours cover the vendor's customer mix, not necessarily yours. Buying does not remove validation. It turns build-and-discover into evaluate-and-verify, which means putting the sensor on your own platform, with your own camera mounting, on your own worst-case sites. That costs far less than a year of in-house tuning.
Staying Competitive After Launch
Shipping does not end the cost. A build commits the team to maintaining the stack for the life of the product line. New constellations and bands arrive, correction services change, image sensors and receivers reach end of life, middleware versions move on, and the research baseline keeps advancing. Positioning technology is not standing still, and a stack frozen at launch ages at the same rate.
Buying moves that burden without removing it. The integrator inherits the vendor's roadmap, which helps when it is ahead of the integrator's and hurts when requirements diverge. It also inherits dependency risk, such as pricing changes, product-line changes and firmware that cannot be inspected. A fair comparison counts compute, correction services, maintenance and after-sales support on both sides of the ledger [9].
Most mitigations are interface decisions:
- Keep the application layer decoupled from the positioning source behind a standard message set.
- Keep wheel odometry as independent aiding.
- Prefer sensors with open drivers. Fixposition publishes open-source ROS 1 and ROS 2 drivers for its visual-inertial GNSS sensors, which lowers the cost of a later swap [10].
If lock-in is the worry, the software-licence route moves the dependency from hardware to an algorithm licence. That is a different risk, though not necessarily a smaller one.
Where the Line Falls
The decision is really two decisions, taken in order. The first is whether the work areas need VIO at all. For open-sky properties an RTK, IMU and wheel stack is a sound answer, and a camera-fusion sensor insures against a failure that will rarely occur. For canopy, building edges and typical residential plots, the question becomes real.
The second is whether you can build it as well as you can buy it, and here the honest answer for most teams is no. A home-built stack is unlikely to match the performance of the PBX-A1, which we regard as the industry-leading option for this class of problem. The reasons follow from everything above. The PBX-A1 is the output of a dedicated fusion team whose only job is this problem, running on calibrated hardware, with confidence reporting built in and a fusion engine that has been tuned against far more platforms and conditions than a single mower programme will see. An in-house team can reach a stack that works. Reaching one that works as well in the hedgerow case, the low-sun case and the wet-lens case takes years of field data, and the vendor is already on that road with a head start.
There are still reasons to build. Do it if positioning is your differentiator and you will staff it as one, if your volume can carry a dedicated team for the life of the product, or if you already hold fleet data a vendor cannot match. Even then, expect the first generation to trail the commercial benchmark and plan the schedule around that. If the aim is to own the algorithms, xFusion Software gets part of the way there without starting from a blank page. Buy if positioning is an input to a product whose value lies elsewhere, if time to market matters, or if engineering capacity is better spent on features customers can see.
The cheapest hedge is to do both on purpose. Fit a PBX-A1 as the reference on the target platform from the first prototype. It sets the performance bar your own work has to clear, and it shows early and cheaply whether an in-house build is closing the gap or never will. Teams that do this often have their answer after one season of data. Teams that skip it tend to reach the same answer after launch, at higher cost.
For help scoping a mower or other outdoor-robot positioning stack, Request a Quote.
References
- MowScout. RTK vs LiDAR vs Vision: How Robot Mower Navigation Works. MowScout (2026). https://mowscout.com/guides/rtk-vs-lidar-vs-vision
- iTechGuides. How Camera Robot Mowers Navigate Without Boundary Wire or GPS. iTechGuides (2026). https://www.itechguides.com/how-camera-based-robot-mowers-navigate-without-boundary-wire-and-sometimes-without-gps/
- Fixposition. OEM Solution: GNSS, Vision & IMU deep fused software, sensor, system-on-module and tailored solution. Fixposition. https://www.fixposition.com/pages/oem-solutions
- xyHt. Beyond the Science Project: Productizing the Future of Sensor Fusion (Q&A with Fixposition). xyHt Magazine. https://www.xyht.com/magazine/magazine-magazine/beyond-the-science-project-productizing-the-future-of-sensor-fusion/
- Fixposition. PBx-A1 documentation: datasheet and integration manual. Fixposition Docs. https://docs.fixposition.com/fd/pbx-a1
- RISE-VIO: Robust Initialization and Targeted Pose Robustification for INS-Centric Visual–Inertial Odometry Under Degraded Visual Conditions. Sensors (2026). https://doi.org/10.3390/s26082305
- Enhancing Visual–Inertial Odometry Robustness and Accuracy in Challenging Environments. Robotics 14, 71 (2025). https://www.researchgate.net/publication/392207717_Enhancing_Visual-Inertial_Odometry_Robustness_and_Accuracy_in_Challenging_Environments
- Greenzie. Greenzie releases 5.0 software, advancing commercial-ready autonomy for the landscaping industry. PRNewswire (10 March 2026). https://finance.yahoo.com/news/greenzie-releases-5-0-software-110200707.html
- Altverse. Robot Lawn Mower Navigation Technologies: Complete Guide. Altverse. https://www.altversebot.com/blog/robot-lawn-mower-navigation-technologies/
- Fixposition. fixposition_driver: ROS driver for Fixposition visual inertial GNSS positioning sensors. GitHub. https://github.com/fixposition/fixposition_driver