Improving Controller Accuracy: Increasing Localization Accuracy Using Sensor Fusion William Chastek, Goktug Poyrazoglu, Yukang Cao, Volkan Isler Abstract Accurate mobile robot localization enables safe and efficient navigation. In this report, the impact of using multiple sensors for local- ization was studied. Two sensors configura- tions were tested: IMU and IMU + LiDAR. This was deployed on both a simulated Turtle- bot3 and a physical F1TENTH robotic plat- form, and was tasked to navigate to a goal po- sition using a logMPPI controller. For both the simulation and real-world cases, the effects of using multiple sensors on the controller ac- curacy were studied. Accuracy was measured by calculating the Euclidean distance between the readings given from the robot versus the ground truth. This was accomplished by track- ing the robot model in Gazebo, and using a PhaseSpace system for the physical robot. Using these metrics it was found that in an idealistic environment where sensor noise is minimal, sensor fusion is not the optimal op- tion. However, in real-world scenarios, it was demonstrated that a particle filter fusing IMU and LiDAR readings reduced average localiza- tion error by at least 11%. This means real- world robotic platforms should aim to use mul- tiple sensors for localization, to increase accu- racy. 1 Background and Motivation Mobile robot localization, the process of estimating a robot’s pose within an environment, underpins all higher-level autonomous tasks such as mapping, path planning, and control. Accurate localization enables safe and efficient navigation, making it a cornerstone of autonomous systems research. Sensor fusion com- bines information from different modalities to exploit complementary strengths: Inertial Measurement Units (IMUs) provide high-rate motion estimates but suffer drift, and Light Detection and Ranging (LiDAR) offers precise range measurements with lower update rates. Prior works demonstrate that fusing multiple sensors can significantly reduce localization error compared to single-sensor setups (Huang et al., 2025)(Cadena et al., 2016), but do not test on physical robotic systems. Recent advances in tightly-coupled fusion algo- rithms have shown error reductions in benchmark sim- ulations (Liu et al., 2022), yet these gains often di- minish when deployed on real hardware due to un- modeled noise and dynamics. Moreover, simulation environments typically simplify sensor artifacts and robot kinematics, creating a sim-to-real gap that must be quantified before field deployment (Abou-Chakra et al., 2025). By evaluating identical fusion and non- fusion pipelines on both simulated and physical robots, this work aims to characterize how model assumptions translate to real-world performance. The results will inform guidelines for the design of robust localization stacks suitable for both development in simulation and reliable operation in the field. 2 Methodology There are two different sensor configurations that were tested, 1. IMU only: Inertial Measurement Unit data alone 2. IMU + LiDAR: Combining IMU data with Li- DAR measurements. For sensor fusion, a particle filter (Künsch, 2013) was developed. The particle filter takes in both the IMU and LiDAR reading, and updates the particles posi- tions based on the IMU readings. Each particle has an (x, y, theta) value, and the particle filter takes the Li- DAR scan and overlays it with each particle. The par- ticles are then weighted based on how close they are to the expected reading of the LiDAR in the map at each position. Lower weighted particles are removed, and new particles are initialized around the particle with the highest weight. For experimentation, there are two main parts: 1. Simulated experiments using the Turtlebot3 pack- age in Gazebo Classic 2. Real-World experiments using a 1:10 sized F1TENTH robotic platform (O’Kelly et al.) Simulated Environment For the simulation portion of the experiments, a subsection of the BARN dataset (Perille et al., 2020) was used. The BARN (Benchmark Autonomous Robot Navigation) dataset comprises 300 Gazebo worlds, each with corresponding maps, occu- pancy grids, and difficulty metrics. Specifically, worlds 0, 198, and 250 were selected out of the 300 worlds uniformly at random. The robot used was Turtle- bot3(Waffle), and it was deployed each time with the same pose and the same goal position for each world. The robot’s ground truth location was tracked using Gazebo’s model, tracking service. Figure 1: An example of a BARN Gazebo world Real-World A physical environment was con- structed and mapped using a ROS SLAM package, slam_toolbox (Macenski and Jambrecic, 2021). The robot was deployed with the same pose each time and the same goal position for each run of the experiment. The robot’s real location was tracked using a PhaseS- pace system (an optical motion-tracking system using LED markers (PhaseSpace Inc., 2023)), which tracked LED markers placed on the robot as it navigated this environment. In both scenarios, the robots were deployed and tasked with navigating to a goal position using a logMPPI con- troller (Mohamed et al., 2022). The robot’s ground truth location, and its belief location were tracked. The difference between the belief and truth positions was measured using Euclidean distance and plotted. 3 Analysis and Results As depicted in Figure 2, in an idealistic environment, probability based localization systems cannot compare to wheel encoders or IMU sensors. After 15 runs, 5 per selected world, the IMU was found to have an av- erage error of 0.0003m1 compared to the ground truth. Conversely, the particle filter has an average error of 0.22m compared to the ground truth. The accuracy of the particle filter based localization is lower when compared to the raw IMU localization because the par- ticle filter expects the IMU and LiDAR readings to have some noise. In noise-free simulations, IMU drift 1Due to Gazebo’s noise-limited IMU simulation, drift was negligible. Table 1: Localization Error Comparison Configuration Simulation (m) Real-World (m) IMU-only 0.0003 0.67 IMU+LiDAR 0.22 0.59 is negligible, while particle filters introduce computa- tional approximations (e.g., finite particles, discretized maps) that dominate error. When implemented, and de- Figure 2: Visualization of the robot trajectories in world 250 from the BARN dataset. The odometry read- ing and the tracked model overlap with each other. ployed on a real-world robot, as depicted in Figure 3, and Table 1 this advantage is lost. After 5 runs with the same goal and start position, the IMU based odom- etry was found to have an average distance of 0.67m compared to the ground truth. Meanwhile, the parti- cle filter has an average distance 0.59m compared to the ground truth. From Figure 3 the error of the par- ticle filter is high when compared to the results found in simulation, about 0.3m of offset. The 0.3m offset could be explained from the robot’s velocity (1m/s) di- vided by the particle filter’s update rate (3 Hz), yielding ∼0.33m per update cycle. This latency was not sub- tracted from reported errors. Another aspect of the real-robot system that showcases the strength of sensor fusion is compounding sensor error. As depicted in 4, while the error rate of the IMU based odometry con- tinues to increase, the particle filter’s error rate stays stable. This means that if you plan to deploy robots for a long period, sensor fusion is a good way to mitigate this accumulating error. Figure 3: Both robot particle filter estimate as well as odometry readings, plotted with the tracked model po- sition using the PhaseSpace system. Figure 4: The average error of each localization method plotted against time. It can be observed that the IMU based localization has accumulating error, while the particle filter based localization remains stable over time. 4 Discussion 4.1 Limitations In this experiment only two sensor configurations were tested. Future works should use more sensor configu- rations, such as depth-cameras, GPS, and/or other sen- sors that can be fitted on a robot. The particle filter’s 3 Hz update rate introduced a latency-derived offset of ∼0.33m (1m/s / 3 Hz). While this explains part of the error, it underscores the need for faster hard- ware or algorithmic optimizations like GPU accelera- tion. GPU-accelerated ray casting could improve par- ticle filter update rates by parallelizing LiDAR scan matching, which would minimize this error. 5 Conclusion In this work, a particle filter-based localization sys- tem that combines IMU and LiDAR measurements was implemented and evaluated on the F1TENTH robotic platform. In simulation, our LiDAR+IMU fusion achieved an average position error of 0.22m (versus 0.0003m for the idealized IMU-only case), and on the real robot reduced mean localization error from 0.67m to 0.59m — a 11% improvement in noisy environ- ments. Our particle filter maintained stable, bounded error over extended runs (see Fig. 4). These results demonstrate that even a minimal sen- sor suite of IMU and LiDAR can significantly mitigate odometric drift in both idealized and practical settings. However, our approach incurs scan processing latency that was not subtracted from the error metrics, and per- formance can still degrade under heavy dynamic obsta- cles or poor LiDAR returns. In addition, the param- eter adjustment for the process and measurement co- variances remains manual and environment specific. Future work will explore adaptive noise estimation to reduce manual calibration, integration of vision- based features for richer landmark constraints, and a closed-loop control evaluation to quantify end-to-end path-tracking gains. We also plan to open source our filter implementation and the synchronized sensor dataset to facilitate reproducibility and further research in low-cost sensor fusion for autonomous robotics. 6 Code and Acknowledgments This project was supported by the University of Min- nesota’s Office of Undergraduate Research. All code used in this project can be found in this GitHub repos- itory, https://github.com/tubbleWeek/ WCHAS_UROP_2025. 7 Hyper parameters For both the simulation and real-world experiments, these hyper parameters were kept constant. 1. Number of Particles: 100 2. LiDAR Sensor Model: https://github.com/tubbleWeek/WCHAS_UROP_2025 https://github.com/tubbleWeek/WCHAS_UROP_2025 • z-hit: 0.85 • z-rand: 0.15 • sigma-hit: 0.1 3. Motion Model: • alpha1: 0.1 • alpha2: 0.1 • alpha3: 0.1 • alpha4: 0.1 alpha1 and alpha3 are the expected translational noise in the robot’s IMU readings, while alpha2 and alpha4 are the expected rotational noise in the robot’s IMU readings. References Jad Abou-Chakra, Lingfeng Sun, Krishan Rana, Bran- don May, Karl Schmeckpeper, Maria Vittoria Min- niti, and Laura Herlant. 2025. Real-is-sim: Bridging the sim-to-real gap with a dynamic digital twin for real-world robot policy evaluation. Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, Jose Neira, Ian Reid, and John J. Leonard. 2016. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on Robotics, 32(6):1309–1332. Zexia Huang, Guoyang Ye, Pu Yang, and Wanshun Yu. 2025. Application of multi-sensor fusion localiza- tion algorithm based on recurrent neural networks. Hans R. Künsch. 2013. Particle filters. Bernoulli, 19(4):1391–1403. Rui Liu, Jun Xu, Yunjiang Lou, and Haoyao Chen. 2022. Multi-sensor fusion based indoor mobile robot localization. In 2022 IEEE 18th International Conference on Automation Science and Engineering (CASE), pages 22–27. Steve Macenski and Ivona Jambrecic. 2021. Slam tool- box: Slam for the dynamic world. Journal of Open Source Software, 6(61):2783. Ihab S. Mohamed, Kai Yin, and Lantao Liu. 2022. Autonomous navigation of agvs in unknown clut- tered environments: Log-mppi control strategy. IEEE Robotics and Automation Letters, 7(4):10240– 10247. Matthew O’Kelly, Hongrui Zheng, Dhruv Karthik, and Rahul Mangharam. F1tenth: An open-source eval- uation environment for continuous control and rein- forcement learning. Proceedings of Machine Learn- ing Research, 123. Daniel Perille, Abigail Truong, Xuesu Xiao, and Pe- ter Stone. 2020. Benchmarking metric ground nav- igation. In 2020 IEEE International Symposium on Safety, Security and Rescue Robotics (SSRR). IEEE. PhaseSpace Inc. 2023. Phasespace IMPULSE motion capture system. https://www.phasespace. com/. Technical specifications: 240Hz tracking, sub-millimeter accuracy, LED marker-based system. http://arxiv.org/abs/2504.03597 http://arxiv.org/abs/2504.03597 http://arxiv.org/abs/2504.03597 https://doi.org/10.1109/tro.2016.2624754 https://doi.org/10.1109/tro.2016.2624754 https://doi.org/10.1109/tro.2016.2624754 https://www.nature.com/articles/s41598-025-90492-4?utm_source=chatgpt.com#Abs1 https://www.nature.com/articles/s41598-025-90492-4?utm_source=chatgpt.com#Abs1 https://doi.org/10.3150/12-BEJSP07 https://doi.org/10.1109/CASE49997.2022.9926435 https://doi.org/10.1109/CASE49997.2022.9926435 https://doi.org/10.21105/joss.02783 https://doi.org/10.21105/joss.02783 https://doi.org/10.1109/LRA.2022.3192772 https://doi.org/10.1109/LRA.2022.3192772 https://par.nsf.gov/biblio/10221872 https://par.nsf.gov/biblio/10221872 https://par.nsf.gov/biblio/10221872 https://www.phasespace.com/ https://www.phasespace.com/ Background and Motivation Methodology Analysis and Results Discussion Limitations Conclusion Code and Acknowledgments Hyper parameters