Two New Sensor Fusion Papers Show How Far We Still Have to Go on Robot Navigation
A pair of arXiv papers tackle the same fundamental problem from different angles, and the results reveal just how much room for improvement remains in autonomous vehicle localization.
Image credit: Lottie animation by Centre Robotics (LottieFiles Free, used with credit). · source
If you've ever watched your phone's GPS struggle in a parking garage, you already understand the core challenge facing autonomous robots: no single sensor tells the whole truth. GPS loses signal. IMUs drift. Wheel encoders slip. The art of sensor fusion is making these unreliable narrators agree on a coherent story, and two recent arXiv papers illustrate just how much work remains in this space.
The papers, both posted this month, attack the same fundamental problem from different angles. One proposes a clever trick to squeeze more information out of existing GNSS data. The other offers an open-source ROS 2 package that tries to fuse everything at once. Neither is revolutionary, but together they paint an interesting picture of where navigation research stands in 2025.
The first paper, Enhanced INS/GNSS State Estimation using GNSS-Based Acceleration Measurements, tackles a specific limitation in standard INS/GNSS fusion. To be precise, the issue is observability: when a vehicle moves slowly or in straight lines, position-only GNSS updates don't give the filter enough information to correct orientation errors or estimate inertial sensor biases accurately.
The proposed solution is to derive acceleration measurements from past GNSS positions using a motion model, then feed those back into the filter as additional observations. It's an elegant idea, actually, because it extracts more signal from data you're already collecting. The authors report mean position RMSE improvements of 11.40% and 20.74% on two unmanned ground vehicle datasets.
Related coverage
More in Autonomy
New research shows convex-guided neural sampling can cut robot path planning time by up to 98%, though the real-world implications remain murky.
Mark Kowalski · 1 hour ago · 5 min
While everyone's chasing humanoid hype, two new papers tackle the unsexy problem of knowing where your robot actually is.
Mark Kowalski · 1 hour ago · 5 min
Two new papers show AI-guided planners can cut computation time by up to 98% while finding better paths. The catch? We're still figuring out when that actually matters.
Sarah Williams · 1 hour ago · 4 min
Two new papers tackle robot learning from opposite ends, and both reveal why the field's been stuck in circles for years.
The second paper introduces FusionCore, a 23-state Unscented Kalman Filter that fuses IMU, wheel encoders, GPS, and Visual SLAM into a single 100 Hz odometry stream. The 23rd state is particularly interesting: an online estimate of systematic yaw rate bias in the wheel encoder, which the filter subtracts during GPS blackouts to reduce heading drift.
FusionCore's headline claim is strong performance against robot_localization, a widely-used ROS package, on the NCLT dataset. The authors report lower Absolute Trajectory Error on 10 of 12 sequences, with improvements ranging from 1.2x to 22.2x. They also note that robot_localization's UKF diverged numerically on all twelve sequences.
I should be clear about what these papers represent. Neither is a paradigm-shattering breakthrough. Both are solid, incremental contributions to a well-studied problem. This isn't a criticism (I know I'm being picky here, but I think it's worth noting that most useful research is incremental).
The GNSS acceleration paper builds on decades of work in INS/GNSS integration. The core insight, that you can derive velocity and acceleration from position histories, isn't new. What's useful is the specific implementation and the empirical validation on real UGV data. The 11-21% improvements are meaningful for practical applications, though the paper doesn't disclose the absolute error magnitudes, which makes it harder to assess real-world significance.
FusionCore is more ambitious in scope but raises some methodological questions:
The comparison against robot_localization is useful, but the claim that it "diverges numerically on all twelve sequences" is striking. It would be helpful to understand why, and whether this reflects fundamental limitations or tuning issues.
The NCLT dataset is from 2012-2013. While it's a standard benchmark, it's worth asking whether results transfer to modern sensor configurations.
The paper doesn't include comparisons against other state-of-the-art fusion approaches, only robot_localization. This makes it difficult to contextualize the results.
The 22.2x improvement on the best sequence is impressive, but the 1.2x improvement on the worst winning sequence suggests highly variable performance.
None of this invalidates the work. The code is open-source under Apache 2.0, which means others can evaluate and extend it. That's genuinely valuable.
Sensor fusion might seem like a solved problem to outsiders. It isn't. Every autonomous vehicle company, every mobile robot startup, every drone manufacturer struggles with localization. The gap between "works in the lab" and "works reliably in deployment" remains substantial.
The GNSS acceleration approach addresses a specific failure mode: performance degradation during low-dynamic motion. This matters for applications like agricultural robots, which spend long periods moving slowly in straight lines, or last-mile delivery vehicles navigating parking lots.
FusionCore's contribution is different. By providing a well-documented, open-source implementation that handles multiple sensor modalities, it lowers the barrier for researchers and small teams who don't want to implement their own fusion stack from scratch. The automatic VSLAM recovery feature is particularly practical; visual odometry systems fail regularly, and graceful degradation is essential.
The yaw bias estimation is interesting from a research perspective. Wheel encoder drift is a known problem, but explicitly modeling systematic bias as a filter state (rather than treating it as noise) is a reasonable approach. Whether the cross-covariance method for identifying this bias generalizes to other platforms remains unclear.
Both papers leave significant questions unanswered. For the GNSS acceleration work:
How does the approach perform with consumer-grade GNSS receivers versus the RTK systems common in research?
What happens in urban canyons or under tree cover, where GNSS quality degrades?
The motion model assumes certain vehicle dynamics. How sensitive are the results to model mismatch?
For FusionCore:
The 23-state filter is computationally heavier than simpler approaches. What are the real-time performance characteristics on embedded hardware?
How does performance compare when VSLAM is unavailable entirely, not just temporarily lost?
The Mahalanobis gating for outlier rejection is calibrated to measurement degrees of freedom. How robust is this to non-Gaussian noise, which is common in real sensors?
I'd also want to see both approaches tested on more recent datasets with modern sensor suites. The field has moved toward tighter integration of learning-based perception with classical state estimation, and it's not clear how these methods interact with, say, learned visual odometry systems.
If I were advising students in this space (and I'm not, to be clear, just a former academic with opinions), I'd push for a few things:
First, rigorous ablation studies. FusionCore has many components: the 23rd state, the adaptive noise covariance, the outlier gating, the ECEF handling. Which ones actually matter? A paper that systematically disables each feature and measures the impact would be more informative than aggregate performance numbers.
Second, failure mode analysis. Both papers report mean improvements, but means hide a lot. What do the failure cases look like? When does the GNSS acceleration approach make things worse? When does FusionCore's yaw bias estimate diverge?
Third, computational benchmarks. Academic papers often run on desktop machines with ample resources. Real robots have power and compute constraints. A comparison of accuracy versus computational cost would help practitioners make informed choices.
Finally, longer-term evaluation. The NCLT sequences are 55-92 minutes each, which is good. But many robotics applications involve hours or days of continuous operation. How do these methods handle long-term drift? Do the bias estimates remain stable?
Sensor fusion research can feel like it's stuck in a local optimum. We keep refining Kalman filter variants, adding states, tuning parameters, but the fundamental approach hasn't changed dramatically in decades. Meanwhile, learning-based methods are making inroads in perception, planning, and control.
There's an open question about whether the future of localization is more sophisticated model-based fusion or end-to-end learned approaches. Probably some combination, but it's too early to say which aspects of classical fusion will survive and which will be subsumed by neural networks.
For now, papers like these represent the working edge of practical robotics. They're not flashy. They won't make headlines. But they address real problems that real systems face, and they do so with enough rigor to be useful. That's actually the research shows what most robotics progress looks like: small steps, carefully validated, building on decades of prior work.
The code for FusionCore is available on GitHub. The GNSS acceleration paper doesn't mention a code release, which is unfortunate. Reproducibility in robotics research remains inconsistent, and I'd encourage the authors to consider making their implementation available.