Resilient and robust odometry is crucial for autonomous systems operating in complex and dynamic environments. Existing odometry systems often struggle with severe sensory degradations and extreme conditions such as smoke, sandstorms, snow, or low-light conditions, threatening both the safety and functionality of robots. To address these challenges, we present Super Odometry, a sensor fusion framework that dynamically adapts to varying levels of environmental degradation. Super Odometry employs a hierarchical structure to integrate four core modules from lower-level to higher-level adaptability including adaptive feature selection, adaptive state direction selection, adaptive engine selection, and a novel learning- based inertial odometry. The inertial odometry, trained on over 100 hours of heterogeneous robotic platforms, captures comprehensive motion dynamics. Super Odometry elevates the inertial measurement unit (IMU) to equal importance with camera and LiDAR within the sensor fusion framework, providing a reliable fallback when exteroceptive sensors fail. Super Odometry has been validated across 200 kilometers and 800 operational hours on a fleet of aerial, wheeled, and legged robots, under diverse sensor configurations, environmental degradation, and aggressive motion profiles. It marks an important step towards safe and long-term robotic autonomy in all-degraded environments.
Figures & tables
Figure 1: Hierarchical Adaptation on Sensor Fusion in Robots. We propose a hierarchical adaptation strategy that initiates with adaptive feature selection for visual degradation, moves to adaptive state direction for LiDAR degradation, applies adaptive factor (engine) selection for mixed degradation, and employs learning-based IMU odometry during complete degradation.
Figure 2: Evaluation of 13 types of degradation in a single run. The color-coded trajectory depicts our estimated odometry of a legged robot navigating through over 13 complex degradation scenarios. Despite these difficulties, the final endpoint drift was only 20 cm over a total distance of 2,966 meters.
Figure 3: Robust feature selection in low light, low texture, and obscured situations. The ellipses represent the uncertainty of 2D feature landmarks; larger ellipse indicates greater uncertainty. The ellipse orientation reflects the ambiguity in feature tracking along a specific 2D direction.
Figure 4: Illustration of adaptive state direction selection in a long corridor from the abandoned Pittsburgh hospital. (A) Relative weights of visual and LiDAR odometry across different sections of long corridor environments. (B) Map of the long corridor, color-coded by the contributing modality; the purple circle indicates areas where visual odometry is actively integrated into sensor fusion, while the green circle denotes areas dominated by LiDAR odometry. Note: A modality is considered “dominant” if its weight in the sensor fusion process is greater than 50%. (C) Confidence metric of LiDAR odometry over a distance of 668 meters. (D) Relative confidence of visual and LiDAR odometry in the degeneracy direction, demonstrating more reliable scan registration achieved with our method.
Figure 5: Adaptive engine selection for mixed degradation in sensor-drop scenarios. (A) We used cardboard to sequentially block the LiDAR’s side, front, and back views, and eventually completely obstructed both the LiDAR and visual sensor to evaluate the dropout tolerance for 600 seconds. (B) The corresponding confidence metric of LiDAR Odometry System. (C) Odometry accuracy evaluation: the trajectory color encodes APE error (top), and the bottom shows mapping results with yellow indicating the estimated trajectory.
Figure 6: Evaluation on learning-based inertial odometry. (A) Qualitative pose comparison shows our method (last column) achieves the highest accuracy against Ground Truth (dashed gray). (B) Accuracy improves significantly when the IMU pre-trained model is trained on multi-robot data versus single-robot data. (C) Online adaptation on unseen datasets demonstrates our model outperforms SOTA, as supported by a quantitative comparison of ATE over time. Each iteration shown corresponds to 1s of adaptation time (e.g., Iteration 10 represents 10 seconds).
Figure 7: Robust Performance in a Smoke Scenario. (A) Adaptive engine selection between learning-based IMU and LiDAR odometry in challenging environments, with purple and blue trajectories indicating their respective use. (B) Trajectory and map generated using learning-based IMU odometry in heavy smoke. (C) Self-supervised adaptation fine-tunes the IMU model from “SubT car” to “Tartan Drive.” (D) Confidence-driven adaptive fusion combines IMU and LiDAR odometry outputs using soft fusion, avoiding hard switches.
Robot Platform
AI-IMU [ 36 ]
RNIN-VIO [ 34 ]
TLIO [ 37 ]
IMO [ 30 ]
IMU-Pretrain (Ours)
Improvement
ATE ↓
T-RTE ↓
ATE ↓
T-RTE ↓
ATE ↓
T-RTE ↓
ATE ↓
T-RTE ↓
ATE ↓
T-RTE ↓
ATE
T-RTE
Wheeled (Car) [ 16 ]
7.68
3.33
7.82
5.06
8.12
3.73
8.12
3.73
6.17
2.52
↑ 19.63%
↑ 24.36%
Legged (Quadruped) [ 16 ]
3.23
1.60
3.10
1.58
3.61
1.73
3.35
1.64
1.46
0.79
↑ 54.83%
↑ 50.63%
Handheld (Human) [ 60 ]
8.26
4.89
7.62
5.61
6.96
4.82
10.19
5.67
4.32
1.95
↑ 47.69%
↑ 60.05%
Aerial (Drone) [ 65 ]
4.14
1.45
4.32
1.51
3.93
1.40
3.72
1.34
3.32
1.04
↑ 19.81%
↑ 28.97%
Table 1: Quantitative comparison of pose estimation between the IMU pre-trained model and specialized IMU models. Our IMU pre-trained model outperformed different specialized IMU models, achieving an average of 35.5% on ATE and an average of 41% on T-RTE, respectively, compared to the second-best model across various robotic platforms, including wheeled, legged, handheld, and aerial systems.
Method
Geometric Degradation
Mixed Degradation
Average
F01
F02
F03
U01
U02
Cave03
Corridor01
Floor01
Liu * [ 66 , 67 ]
0.307
0.095
0.629
0.122
0.235
0.260
1.454
0.401
0.588
Weitong * [ 70 , 21 ]
0.26
0.096
0.617
0.120
0.222
0.402
1.254
0.577
0.663
Kim * [ 26 , 69 ]
0.331
0.092
0.787
0.123
0.270
0.279
2.100
0.650
3.825
Yibin [ 71 ]
1.060
0.220
0.750
0.470
0.620
9.140
2.990
5.500
4.312
Zhong * [ 72 , 73 ]
1.205
0.695
-
1.175
1.72
2.08
-
-
1.209
Table 2: ATE performance on SubT-MRS [ 39 ] . * denotes incorporation of loop closure. - denotes incomplete trial.
Figure 8: System Overview. (A) A hierarchical adaptation framework that integrates adaptive sensor fusion with a learning-based IMU odometry. The sensor fusion module provides free pose supervision to guide the learning process of the IMU odometry, helping to reduce drift and capture motion patterns. In turn, the IMU odometry offers motion priors and serves as a fail-safe solution. (B) Illustration of adaptive feature selection, adaptive state direction estimation, and adaptive engine selection.
Figure S1: Robust odometry across diverse conditions with various sensors and robots. We tested our method in various scenarios with visual degradation (A-D), geometric degradation (E-H), mixed degradation(I-L), complete degradation (M-P) and other scenarios such as caves, off-road, steep slopes, and featureless deserts (Q-T).
Figure S2: Large-scale mapping results from the DARPA subterranean challenge . (A) The DARPA Final Challenge featured a combination of tunnel, urban, and cave environments (B) Collaborative mapping using both drones and ground vehicles, with different colors representing the maps generated by each vehicle. (C) Reconstructed a mine environment during a 1-hour run, covering an area of 990,000 m2 with high precision. (D) Visual examples of the mine environments are dark, self-similarity, and lack of distinct features.
Figure S3: Illustration of adaptive state selection and self-weight parameter tuning in a multi-floor environment. (A) The legged robot navigates featureless corridors, narrow staircases, overexposed areas, and transitions between indoor and outdoor environments. (B) The system dynamically identifies degraded directions and predicts alignment risks during point cloud registration. (C) Our system self-tunes the weight between LiDAR and visual sensors to optimize fusion performance across confined and open spaces.
Figure S4: Aggressive motion in off-Road environments. (A) High-fidelity large-scale mapping results. (B) Our ground vehicle traversed uneven off-road surfaces, such as rocks, steep slopes, and unseen ditches, over a 1.38-hour run. (C) The vehicle reached a maximum speed of 13 m/s . Despite these challenges, our method was still able to recover a high-quality map.
Figure S5: Unexpected Collision from a Spinning Drone. (A) Accurate mapping results despite unexpected collisions. Our odometry solution maintained precise state estimation during a hardware failure, enabling the drone’s safe autonomous landing. (B) During the flight, our drone unexpectedly lost control, colliding with walls and spinning erratically for 106 seconds. (C) Angular velocity variations over time. (D) Rapid drone movements in featureless environments.
Figure S6: Mapping Results in Snow Environments. (A) Our legged robot navigates steep inclines, slippery surfaces, and staircases, and heavy snow significantly degrades camera systems. (B) Heavy snow leads to degraded image quality for feature detection. (C) The detailed 3D map of staircases. (D) Our method can still recover detailed 3D maps over a distance of 798 meters.
Figure S7: Non-forgetting Characteristics of Our Domain Adaption. We assessed the generalization capabilities of four distinct models across both the same source domain and a target domain. Unlike other models, which exhibit catastrophic forgetting of source domain knowledge following domain adaptation, the LoRA-based approach (last column) effectively mitigates this issue, preserving robust generalization performance. In the visualization, the dashed line represents the ground truth trajectory, while the color-coded lines depict the estimated trajectories.
Robust multi-sensor fusion is essential for reliable autonomy in diverse and degraded environments, where sensor reliability can fluctuate rapidly. Because different modalities fail in distinct ways, effective fusion should adaptively balance complementary cues rather than rely on fixed weighting. This adaptability is particularly important for ego-motion estimation, since accurate updates depend on the consistent integration of complementary sensor information. We propose ADM-Fusion, an end-to-end deep learning based multi-sensor fusion method designed to adapt to environmental changes and sensor degradation. ADM-Fusion employs an adaptive sensor mixture-of-experts framework with content-aware routing to dynamically assign weights to sensor inputs in real time. The system further incorporates separate translation and rotation branches, coupled through a cross-task attention mechanism to preserve task-specific specialization while enabling information sharing. ADM-Fusion is trained on the CARLA-LOC simulated dataset and subsequently fine-tuned on KITTI real-world data, demonstrating effective simulation-to-real transfer. Experiments show that ADM-Fusion remains robust under degraded conditions while maintaining competitive performance against existing methods.
Hasan Moughnieh, Ibrahim Ghaddar, Hadi Elham +2
Department of Electrical and Computer Engineering, American University of Beirut, Beirut, Lebanon. · Department of Mechanical Engineering, American University of Beirut, Beirut, Lebanon.
Robust locomotion and navigation for legged robots relies heavily on dependable odometry. Traditional multi-sensor fusion for such state estimation requires meticulous sensor calibration and platform-specific kinematic modeling, which complicates deployment. Industrially packaged exteroceptive sensors can provide accurate motion tracking but remain vulnerable to perceptually degraded conditions. We thus develop a plug-and-play, robust legged odometry system that eliminates the need for explicit exteroceptive-to-proprioceptive calibration or system kinematic modeling. Our approach leverages established exteroceptive motion pipelines as a continuous supervisory signal to train an online learned velocity neural network directly from proprioceptive data. An Invariant EKF (InEKF) is then used to fuse the learned proprioceptive or exteroceptive velocity (if any) and IMU data. When exteroception fails due to environmental degradation, the system seamlessly falls back to using the learned proprioceptive model, yielding a resilient legged odometry that readily adapts to new hardware. We demonstrate the platform-agnostic, easily deployable nature of our approach on different quadruped platforms, showcasing promising results in maintaining robust motion estimation across challenging scenarios.
Accurate localization in GPS-denied and visually degraded environments affected by smoke, dust, or fog remains a critical challenge for autonomous systems. While LiDAR provides high-resolution spatial perception, its performance deteriorates under particulate occlusion. In contrast, frequency-modulated continuous-wave (FMCW) radar offers robustness under such conditions but suffers from sparse measurements and outliers. This article presents a resilient decentralized LiDAR Radar Inertial Odometry (LRIO) framework that fuses the complementary strengths of LiDAR, radar, and inertial sensing to achieve robust pose estimation in challenging subterranean environments. At the core of the framework is an Adaptive Residual Thresholding (ART) mechanism that autonomously detects temporal sensor degradation and selectively adjusts the fusion process to minimize unreliable measurements. Building on this, a Robust Error-Tolerant Optimal Information Filter (RETOIF) is employed within a decentralized fusion architecture to integrate LiDAR-inertial and multiple radar-inertial odometry estimates in a fault-aware manner, enhancing resilience to environmental disturbances and intermittent sensor failures. The proposed framework is evaluated using real-world sensor data collected in GPS-denied subterranean tunnels under smoke-free and smoke-degraded conditions. In the smoke-free experiment, the proposed RETOIF framework achieves a root mean square localization error (RMSE) of 2.1733, compared with 2.9929 for the conventional centralized LRIO approach, corresponding to a 27.4% reduction in RMSE. Experimental results further demonstrate the framework's ability to maintain pose estimation under smoke-induced sensor degradation, highlighting its potential for resilient autonomous navigation and mapping in challenging subterranean environments.
Moumita Mukherjee, Avijit Banerjee, Anton Koval +1
Robotics and AI Team, Department of Computer Science, Electrical Engineering and Space Engineering, Luleå University of Technology, Luleå, Sweden