A preprint reports that an odometry and sensor-fusion system completed real-time state estimation without failures during a comprehensive degradation run. Endpoint drift was 0.2 metres over 2966 metres, while the largest reported absolute trajectory error was 0.184 metres during a 600-second sensor-drop test.
The broader evaluation reported advantages across four robotic platforms without fine-tuning, and the authors' model's error fell sharply during 60 seconds of online adaptation. The evidence comes from robot runs and benchmarks in particular tested conditions, so it does not establish universal reliability.
A system that changes tactics
The study asks whether the approach can generalize across platforms, adapt online to unseen environments and improve sensor-fusion robustness in extreme settings.
Its design is an escalation ladder: adaptive feature selection comes first, followed by adaptive state-direction selection and adaptive engine selection, with learning-based IMU odometry used as degradation becomes more severe.
The test across robots
Training for the IMU model used more than 100 hours of real-world inertial data, with ground-truth trajectories across eight platforms. The data came from SubT-MRS, TartanDrive, IDOL, Blackbird and UZH.
Reported validation covered 200 kilometres and 800 operational hours on aerial, wheeled and legged robots under varied sensor configurations, environmental degradation and aggressive motion profiles.
Results under pressure
During the online-adaptation series, the authors' model's ATE fell from 32.87 metres at 0 seconds to 1.73 metres at 60 seconds. A state-of-the-art comparison model fell from 34.14 metres to 4.94 metres over the same period. The reported improvement at 60 seconds was 64.98%.
Without fine-tuning, the pre-trained IMU model reportedly outperformed specialized models across four robotic platforms. Average ATE was 35.5% better and time-relative trajectory error, or T-RTE, was 41.0% better than the second-best model.
For a human-handheld sequence, the paper reported reductions of up to 54% in ATE, 117% in T-RTE and 100% in distance-relative trajectory error, or D-RTE, when diverse robot training was compared with human-handheld-only training. The paper does not explain how the 117% T-RTE reduction was calculated.
Across robots and benchmarks
On eight SubT-MRS sequences, the reported real-time average ATE was 0.271, described as 54% better than the second-best method. On the robustness benchmark, the reported average scores were 0.925 for position robustness and 0.940 for rotation robustness.
In a smoke demonstration, repeated figure-eight transitions between smoky and clear areas triggered adaptive switching and maintained robust state estimation. The reconstructed smoke map was severely noisy, and no quantitative endpoint error was reported for that test.
What the results do not settle
Metrics were computed with Python and NumPy, and all trajectories were upsampled to 200 hertz for comparison. The evaluation reported no p-values, confidence intervals or formal inferential tests, and several headline demonstrations were single runs or qualitative evaluations.
Complete-degradation and smoke-related cases were excluded from benchmark averages because baseline algorithms failed under those conditions. The benchmark comparisons also need context: competitor results came from the open challenge, and competing teams could use loop closure and offline processing.
The reported numbers describe performance on the tested platforms, datasets and conditions, not universal reliability. The system still depends on accurate calibration and time synchronization, requires manual parameter tuning, and has distribution gaps on unseen robots and environments; faster adaptation and lower computation and memory demands for large underground deployments remain open needs.
The evidence trail
The preprint says all data needed to evaluate its conclusions are in the paper or Supplementary Materials, with study data deposited in Zenodo. It lists Supplementary Figures S1 to S7 and Movies S1 to S6.
Funding was reported from the U.S. Army Research Lab under W911NF-23-S-0001, W911NF2120152, W911NF2420125 and W911NF-17-S-0003. The preprint is listed as arXiv:2608.25427v1, dated 26 August 2026.
Paper data and sources
Original title: SUPER ODOMETRY 2.0: Resilient Odometry via Hierarchical Adaptation
Authors: Shibo Zhao, Sifan Zhou, Yuchen Zhang et al.
Journal/Repository: Science Robotics 2025
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: 10.1126/scirobotics.adv1818
Original paper · Full text