The new framework, called MSENet, short for Multi-Stage Extraction Network, was described in a study published in the journal Mobile Networks and Applications. The team, led by Gan Yuxi of Tsinghua University and including collaborators from Guangxi University of Science and Technology, Shenzhen Jinzhi Technology, Xinjingjie Digital Health, the Shenzhen Institute of Information Technology, and WeiKang Intelligence Technology, set out to solve a problem that has become acute with the explosion of the Internet of Things, mobile robotics, and wireless sensor networks: how to deliver high-accuracy inertial positioning with low latency and low power consumption on hardware that fits in a shoe, a drone, or a palm-sized robot.

The core insight behind MSENet lies in how it extracts information from raw IMU signals. Instead of relying on a single, monolithic feature-extraction pipeline, the network employs a multi-stage strategy in which motion representations are built up progressively across several processing stages. Each stage refines the temporal understanding of the motion captured by the accelerometer and gyroscope, allowing the model to capture both fine-grained, short-term fluctuations and longer-term patterns of human or robotic movement. According to the authors, this design enhances the temporal motion representation while simultaneously reducing computational overhead, because the early stages can use lightweight operations to filter and condense the signal before more expressive layers operate on a more compact representation. The approach echoes lessons learned in efficient computer-vision architectures such as MobileNetV2, MnasNet and EfficientNet, which the authors cite as inspiration, but adapts them to the peculiarities of inertial data — continuous, high-frequency waveforms rather than discrete images.

The benchmarks the team chose are the de facto standards of the field. RIDI, or Robust IMU Double Integration, introduced by researchers at the European Conference on Computer Vision in 2018, tests how well a model can reconstruct a trajectory from raw inertial readings alone. RONIN, short for Robust Neural Inertial Navigation in the Wild, presented at the IEEE International Conference on Robotics and Automation in 2020, pushes the challenge further with real-world pedestrian data collected in diverse environments. On both benchmarks, MSENet achieves accuracy comparable to IMUNet, an efficient regression architecture published in IEEE Transactions on Instrumentation and Measurement in 2024 that represents the current state of the art. But the comparison tells only half the story. MSENet achieves this parity while reducing the number of model parameters by 69.95 percent and the number of floating-point operations, or FLOPs, by 44.75 percent — a reduction so large that it changes what kind of hardware the model can realistically run on.

Perhaps the more impressive result comes from the RONIN-UNSEEN dataset, a split of the benchmark containing environments and motion patterns the network never encountered during training. Generalization — the ability to perform well on data unlike the training set — has been a persistent weakness of learned inertial navigation systems, which often overfit to the specific gait patterns, sensor orientations and building layouts they were trained on. On this challenging unseen set, MSENet reduces the Absolute Trajectory Error by 11.5 percent and the Relative Trajectory Error by 4.3 percent compared with state-of-the-art methods. In practical terms, absolute trajectory error measures how far the estimated final position deviates from the true position over an entire walk or route, while relative trajectory error captures local accuracy over shorter segments. A model that is both smaller and more generalizable is a rare combination in deep learning, where efficiency gains typically come at the cost of robustness.

The significance of these numbers becomes clear when one considers where inertial navigation must actually live. The accelerometers and gyroscopes in a smartphone, a smartwatch, a wearable medical monitor or an autonomous warehouse robot are governed by strict power and thermal budgets. A deep network with millions of parameters may achieve impressive accuracy on a laboratory GPU, but it cannot run in real time on a microcontroller drawing milliwatts. Earlier efforts to address this include TinyOdom, a hardware-aware neural inertial navigation system published in the Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies in 2022, which demonstrated that aggressive efficiency constraints could be built into neural navigation from the ground up. MSENet extends this line of thinking with a multi-stage extraction design that squeezes more representational power out of every parameter.

The problem the researchers are attacking is as old as inertial navigation itself. Traditional systems rely on Kalman filtering and other probabilistic estimation techniques, often fusing IMU data with auxiliary sensors such as shoe-mounted detectors, radio signals or camera feeds to correct the inevitable drift. Landmark work in the mid-2000s, including Foxlin’s pedestrian tracking system with shoe-mounted inertial sensors, showed how zero-velocity updates — brief moments when the foot is planted and the sensor is momentarily still — could reset accumulated error. But these methods require careful engineering, specialized hardware placement or external references, and they degrade when those aids are unavailable. Learned approaches such as IONet, presented at the AAAI Conference on Artificial Intelligence in 2018, and later RONIN and IMUNet, took a different route: train a neural network end to end to map raw IMU waveforms directly to position or velocity, letting the model learn to recognize and compensate for sensor noise, bias and drift on its own. The price of that flexibility has been computational weight, and MSENet is an attempt to pay less of it.

The applications extend well beyond counting steps. In dense urban canyons where satellite signals bounce off buildings, in underground transit systems, in disaster-response scenarios where GPS is jammed or absent, and in the coordinated fleets of autonomous mobile robots now populating factories and hospitals, reliable inertial positioning is the silent backbone of every navigation stack. The authors frame their work explicitly in the context of wireless sensing and the Internet of Things, noting that inertial localization is increasingly a component of wireless sensor networks where edge devices must compute locally to reduce latency, preserve bandwidth and protect privacy. A navigation model small enough to run on the sensor node itself, rather than shipping raw data to a server, aligns with a broader industry trend toward edge intelligence — the migration of machine-learning inference from the cloud to the point of data collection.

The research was supported in part by the National Key Research and Development Program of China, under Grant No. 2025YFE0103200, reflecting the strategic weight that national funding programs now place on autonomous navigation and sensing technologies. The author team spans academia and industry in a pattern that has become characteristic of Chinese applied AI research: Gan Yuxi and Xia Yin from Tsinghua University’s Department of Computer Science and Technology, Jiongxun Lin from Guangxi University of Science and Technology, Zhihu Jin and Xiao Chen from Shenzhen Jinzhi Technology, Daoquan Huang from Xinjingjie Digital Health, and Bao Peng from the Shenzhen Institute of Information Technology. According to the published contribution statement, Gan Yuxi conceived the study, designed the MSENet framework, implemented the algorithms and led the experiments, with the co-authors dividing responsibilities across data processing, validation, performance evaluation and supervision. Correspondence for the study is handled by Jiongxun Lin.

The paper was received on 5 June 2026, accepted on 2 September 2026, and published in the journal’s 11 September 2026 issue under the exclusive license of Springer Science and Business Media. The article’s data availability statement notes that no new datasets were generated or analysed beyond the public benchmarks used in evaluation, meaning the headline results — the 69.95 percent parameter reduction, the 44.75 percent FLOP reduction, the 11.5 percent improvement in absolute trajectory error — can be independently checked against the widely available RIDI and RONIN data by any laboratory with the model weights and standard evaluation protocols. That reproducibility is a quiet but important feature in a field where benchmark claims have sometimes been difficult to replicate.

What remains to be seen is how far the multi-stage extraction philosophy can scale. The benchmarks cover pedestrian motion, and the authors position the framework for smart robots, wearable technology, motion detection and edge computing more broadly — but robot dynamics, with rapid turns, vibrations and payloads, present distributions of inertial signal quite unlike human gait. The generalization results on RONIN-UNSEEN are encouraging on this front, suggesting the learned temporal representations are not merely memorizing walking patterns. If the same efficiency-accuracy balance holds on robotic platforms and across heterogeneous sensor hardware, MSENet could join a growing toolbox of lightweight neural navigators that make precise, self-contained positioning a default capability of the billions of low-power devices now coming online. For a technology whose central paradox — sensors that get better the smaller and smarter they become — has defined its limits for decades, a 70 percent cut in model size with accuracy to spare is the kind of result that reshapes what engineers dare to attempt on a circuit board the size of a fingernail.