PaperlayerView exact revision on arXiv

Source-linked AI summary

SUPER ODOMETRY 2.0: Resilient Odometry via Hierarchical Adaptation

Shibo Zhao, Sifan Zhou, Yuchen Zhang, Ji Zhang, Chen Wang, Wenshan Wang, Sebastian Scherer

arXiv:2608.25427v1cs.RO

TL;DR

Robust odometry remains difficult under severe, persistent, and varied sensor degradation, while existing systems struggle with generalization and computational efficiency. Super Odometry addresses this gap with hierarchical adaptation, reciprocal learning between inertial and traditional estimators, and learning-based IMU odometry. The framework provides robust state estimation across degraded environments, while calibration, synchronization, and unseen-domain generalization remain limitations.

  • Problem

    Existing odometry frameworks struggle to generalize across diverse degradation scenarios while balancing robustness and computational efficiency.

    Source: Discussion · p. 10

  • Method

    Super Odometry dynamically reconfigures sensor fusion through hierarchical adaptation and reciprocal learning between a learning-based IMU estimator and traditional model-based fusion.

    Source: Discussion · p. 10

  • Results

    The framework offers robust, adaptable, and efficient state estimation across diverse degraded environments and robotic platforms.

    Source: Introduction · p. 2, Discussion · p. 10

  • Takeaways & Limitations

    The system provides a reliable learned-IMU fallback and supports resilient odometry under real-world sensor degradation.

    Source: Discussion · p. 10

    Performance depends on accurate calibration and precise time synchronization, while the learning-based IMU model struggles with unseen domains caused by training–testing distribution gaps.

    Source: Discussion · p. 10

Abstract

from arXiv · show

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.

Summary

The paper presents a resilient odometry system designed to adapt seamlessly across diverse environments and robotic platforms.

  • Super Odometry is designed as a resilient odometry system that adapts seamlessly to diverse environments and robotic platforms.Source: Summary · p. 1

Introduction

Super Odometry addresses odometry failures under severe, persistent, and varied sensor degradation with a dynamically reconfigurable hierarchical sensor-fusion framework. It combines adaptive responses with learning-based inertial estimation and reciprocal online adaptation, and is evaluated extensively across heterogeneous robots and conditions.

  • Odometry supports robot localization, 3D reconstruction, navigation, exploration, path planning, and control.Source: Introduction · p. 1
  • Existing systems struggle with poor lighting, geometric degradation, motion blur, extreme weather, and simultaneous visual and geometric degradation.These conditions can impair feature tracking and data association, leading to drift or failure.Source: Introduction · p. 1
  • Persistent multi-sensor degradation, high computational cost, and fixed parameter settings remain central challenges for resilient, efficient, and generalizable odometry.The paper highlights smoke affecting visual and LiDAR sensing, costly redundant fusion, and limited transfer across indoor and outdoor environments.Source: Introduction · p. 1
  • The heterogeneous learning-based IMU module predicts 3D poses in real time from internal motion cues and serves as a motion prior when external observations become unreliable.It is trained on diverse robotic data and is designed to maintain low-drift accuracy and generalization.Source: Introduction · pp. 1–2
  • Reciprocal fusion and bilevel self-supervised online adaptation let traditional model-based estimation and the learned inertial network mutually refine one another over time.A lower-level factor graph generates pose labels, while the inertial network supplies motion priors back to the graph.Source: Introduction · p. 2
  • Super Odometry dynamically scales its response from adaptive feature selection through state-direction and engine selection to learning-based inertial odometry.The hierarchy targets mild visual degradation, moderate geometric degradation, mixed degradation, and complete degradation, respectively.Source: Introduction · p. 2, Figure 1 · p. 2
  • Over 200 kilometers and 800 operational hours across aerial, wheeled, legged, and handheld robots, Super Odometry consistently outperformed state-of-the-art techniques under diverse and extreme degradation.The evaluation covered six years of operation and varied sensor and motion conditions.Source: Introduction · p. 2

Results

Super Odometry maintains real-time odometry across visual, geometric, mixed, and complete degradation by adapting feature selection, sensor weighting, engine selection, and learned inertial estimation. Evaluations report low drift, reliable registration, cross-platform generalization, rapid online adaptation, and robust state estimation in smoke.

Visual degradation

  • Adaptive feature selection rejects unreliable visual features under low light, low texture, and obscured conditions.The method estimates feature-correspondence quality and covariance rather than passively tracking all features.Source: Results · p. 3, Figure 3 · p. 4

Geometric and mixed degradation

Learning-based inertial odometry

  • 35.5% lower ATE and 41.0% lower T-RTE than the second-best model were obtained by the heterogeneous IMU pre-trained model across robotic platforms.Training on diverse robot data also reduced ATE by up to 54% in the human-handheld sequence.Source: Learning Based Inertial Odometry for… · p. 9

Online adaptation and smoke

  • Online adaptation converged faster than state-of-the-art methods, while adaptive IMU use maintained robust state estimation during repeated smoke transitions.The smoke experiment repeatedly switched between smoky and clear areas; the reconstructed map remained noisy, but state estimation stayed robust.Source: Learning Based Inertial Odometry for… · pp. 9–10

Discussion

The paper argues that resilient odometry requires hierarchical adaptation rather than rigid fusion, because degradation violates common noise and redundancy assumptions. It further positions learning-based inertial sensing as an independent, mutually adaptive source while identifying calibration, synchronization, and generalization as remaining challenges.

Motivation

  • Rigid multi-modal fusion can fail across diverse degradation scenarios because degraded measurements violate Gaussian-noise and redundancy assumptions.Smoke and darkness can produce corrupted measurements that dominate particular state directions, making estimation unreliable and optimization more expensive.Source: Discussion · p. 10

Hierarchical Adaptation

  • The proposed framework dynamically escalates from lightweight to more complex adaptations as degradation severity increases, covering visual, geometric, mixed, and full sensor failures.This hierarchy is intended to preserve reliable state estimation while matching computational effort to environmental difficulty.Source: Discussion · p. 10

Inertial Sensing

  • Learning-based inertial odometry gives the IMU an independent role alongside LiDAR and visual sensors, rather than limiting it to pre-integration.The paper argues that internal motion sensing can provide state estimation when exteroceptive sensing becomes unreliable.Source: Discussion · p. 10

Reciprocal Fusion

  • Reciprocal fusion lets the inertial model and traditional sensor fusion learn from each other, with the IMU providing motion priors and fallback estimation under degradation.The fusion pipeline adapts the IMU model using feedback while the learned model supports robustness across environments.Source: Discussion · p. 10

Limitations

  • Accurate calibration and precise time synchronization remain important dependencies for system performance.The paper identifies online calibration and time synchronization as directions for reducing manual parameter tuning.Source: Discussion · p. 10
  • The learning-based IMU model still struggles with unseen domains because of distribution gaps between training and testing data.The authors suggest combining real-world and simulated IMU data to improve generalization and adaptation to new robots and environments.Source: Discussion · p. 10

Materials and Methods

Super Odometry combines hierarchical adaptive sensor fusion with learning-based inertial odometry in a self-supervised framework. Its mechanisms adapt estimation to degradation by selecting features, state directions, engines, and learned motion priors.

Hierarchical Adaptation

  • The framework groups adaptive feature selection, state direction, and engine selection as adaptive sensor fusion, alongside separately trained learning-based inertial odometry.Together, these components form the hierarchical adaptation framework.Source: Materials and Methods · pp. 10–11

State Estimation

  • Odometry estimates the agent’s trajectory over time from observations and control inputs without constructing a persistent map.The formulation targets incremental estimation of the next pose.Source: Problem Definition · p. 11

Robustness Evaluation

  • Relative-pose robustness metrics use AUC of F1 scores to assess position and rotation estimation across tolerance levels.Position robustness is Rp = AUC(F1(ve)), while rotation robustness is Rr = AUC(F1(ωe)).Source: Problem Definition · pp. 11–12

Adaptive Sensor Fusion

  • Adaptive state direction analysis estimates LiDAR observability at the front end and adjusts visual weights toward vulnerable state directions.The method predicts ICP alignment risk and computes observability for each point along each state direction.Source: Adaptive Sensor Fusion · p. 13

Learning-based Inertial Odometry

  • Learning-based inertial odometry uses a pretrained model and adapter network, operating on body-frame acceleration and angular velocity to regress body-frame velocity.Body-frame inputs and losses reduce dependence on specific world-frame trajectories and support generalization.Source: Learning-based Inertial Odometry · p. 14
  • LoRA adapts the pretrained IMU model by training low-rank updates while keeping the original weights fixed.The adapter transfers new knowledge while limiting parameter updates and supporting continuous adaptation.Source: Learning-based Inertial Odometry · p. 15

Supplementary materials

The supplementary materials describe Super Odometry’s degradation-aware hierarchy, heterogeneous inertial model, online adaptation, and broad evaluation scope. They position the system as a resilient odometry approach tested across diverse robots and challenging conditions.

Supplementary Materials

System Design

  • Super Odometry reciprocally couples traditional model-based odometry with learned inertial estimation, allowing each component to provide supervision or motion priors to the other.The traditional estimator refines the inertial network under nominal conditions, while the learned module takes over as external sensing degrades.Source: System Overview · p. 19

Heterogeneous Learning-based Inertial Odometry

  • The heterogeneous IMU model uses diverse data from drones, quadrupeds, cars, humans, and other platforms to support cross-platform motion estimation.The model is described as trained on hundreds of hours of varied robotic data.Source: Heterogeneous Learning-based Inertial… · p. 19

Online Adaptation

Evaluation

Adaptive State Direction

Implementation Detail on Learning-based IMU Odometry

The learning-based IMU implementation uses standardized sliding-window inputs, augmentation, and body-frame velocity supervision. Training employs a PyTorch model optimized with Adam over 50 epochs.

Sampling Specifications

  • Each IMU input window contains 10 seconds sampled at 200 Hz, producing an input dimension of 10 × 200 × 6.Lower-frequency data are linearly interpolated to 200 Hz.Source: Implementation Detail on… · p. 22

Data Augmentation

  • Training augmentation adds random Gaussian noise, acceleration and gyroscope biases, and random yaw rotations to each IMU sample.The yaw rotation promotes learning of yaw-angle invariance.Source: Implementation Detail on… · p. 22

Network Setting

  • The model uses relative body-frame velocity as supervision and is trained with Adam at an initial learning rate of 0.0001.Training lasted 50 epochs and approximately 20 hours on an NVIDIA Tesla V100 GPU.Source: Implementation Detail on… · p. 22

DARPA Subterranean Challenge

The system demonstrated robust multi-robot mapping across smoke-filled, dusty, poorly lit subterranean environments and large, repetitive mine structures. It also achieved strong placements across distinct SubT challenge settings.

  • 596.6 m, 499.8 m, 445.2 m, and 112.0 m were covered by multiple robots over 38 minutes in the SubT Final Challenge.The deployment combined aerial drones and ground vehicles in challenging subterranean terrain.Source: DARPA Subterranean Challenge · p. 22
  • Fourth place in the SubT Final Challenge, second in the Urban Challenge, and first in the Tunnel Challenge demonstrated adaptability across environments and robot platforms.Source: DARPA Subterranean Challenge · p. 22
  • 990,000 m2 was reconstructed with high precision during a 1-hour run in a dark, repetitive mine environment.Image and LiDAR place recognition was particularly difficult in this setting.Source: DARPA Subterranean Challenge · p. 22

Adaptive State Direction in Multi-floor Environments

In multi-floor environments, the system adapts state-direction confidence and sensor weights to compensate for weak vertical constraints and changing visual or geometric conditions. It successfully traversed complex indoor-outdoor terrain.

Multi-floor Environments

Adaptive State Direction

Adaptive Weight Parameter Tuning

  • 60% LiDAR confidence outdoors versus 40% visual confidence reflected stronger geometric features and overexposed images.The self-tuning mechanism adjusted sensor contributions according to environmental conditions.Source: Adaptive State Direction in… · p. 23

Aggressive Motion in Off-Road Environments

The system was evaluated under aggressive off-road motion that challenged both visual and LiDAR sensing. It maintained accurate pose recovery and high-fidelity mapping across difficult terrain.

Challenges

Evaluation Conditions

  • 13 m/s maximum speed and approximately 5 m/s2 peak acceleration were sustained during a 1.38-hour off-road run.The vehicle traversed rocks, steep slopes, ditches, and other uneven surfaces.Source: Aggressive Motion in Off-Road… · p. 23

Results

Unexpected Collision from an Off-nominal Spinning Drone

A drone experienced severe locomotion failure after losing a propeller blade, causing uncontrolled rotation and collision. Adaptive sensor fusion maintained useful odometry, reconstructed a 3D map, and enabled safe landing.

Failure Event

Failure Interaction

Recovery

Robustness Experiments in Snow Environments

The experiments evaluate Super Odometry across challenging snowy terrain and domain shifts. The method preserves detailed mapping under heavy snow and uses LoRA-based adaptation to retain source-domain estimation after target-domain adaptation.

Snow-environment mapping

  • Heavy snow degraded camera image quality, while slipping and compressible terrain disrupted locomotion patterns and challenged state estimation.These conditions affected both visual feature detection and motion consistency.Source: Robustness Experiments in Snow… · p. 26
  • 798 meters: the method generated a high-fidelity 3D map across steep slopes, staircases, and indoor–outdoor transitions under heavy snow.The staircase map further demonstrates the system’s reported precision and robustness.Source: Robustness Experiments in Snow… · p. 26

Domain adaptation

  • The pretrained IMU model produced more accurate trajectory estimates than the expert model in the SubT-Dataset source domain.This result is presented as evidence of strong source-domain generalization.Source: Non-forgetting Characteristics of Our… · p. 26
  • In the unseen TartanDrive target domain, the pretrained model without fine-tuning performed poorest, while full fine-tuning produced reasonable trajectories but relatively large pose errors.The passage contrasts these outcomes with the LoRA-based IMU approach, but its remaining result is truncated.Source: Non-forgetting Characteristics of Our… · p. 26
  • After target-domain adaptation, LoRA-based fine-tuning maintained accurate source-domain estimation, whereas fully fine-tuned models exhibited catastrophic forgetting.The comparison covers both RNIN-VIO and the pretrained model with full fine-tuning.Source: Non-forgetting Characteristics of Our… · p. 26

A Mapping result

The figures depict odometry and mapping behavior during aggressive motion, unexpected collision, snow, and domain adaptation. Reported examples include safe landing after a spinning-drone failure and detailed snow-environment mapping over 798 meters.

Aggressive motion

  • 1800 degrees in 3 seconds: the figure labels an aggressive rotational motion profile for the drone.The supplied text provides the rotation magnitude and duration but no associated performance outcome.Source: B Aggressive rotation scenario · p. 27

Unexpected collision

  • 106 seconds: a drone spun erratically after losing control and colliding with walls, while the odometry solution maintained precise state estimation for safe autonomous landing.The figure separates the unexpected collision sequence from the resulting mapping and landing outcome.Source: Figure S5 · p. 27

Aggressive motion

  • Angular velocity is shown over time alongside rapid drone movements in featureless environments.The supplied caption identifies the plotted motion context without reporting a comparison or outcome.Source: C Angular Velocity versus Time · p. 27

Snow mapping

  • 798 meters: the method recovered detailed 3D maps in heavy snow, including staircases and challenging terrain.The figure describes navigation across steep inclines, slippery surfaces, and staircases while snow degraded camera systems.Source: Figure S6 · p. 28

Domain adaptation

  • Fig. S7 compares four models across source and target domains to assess generalization and forgetting after domain adaptation.The LoRA-based approach is described as preserving source-domain generalization better than the other models.Source: Figure S7 · p. 29

Related papers

Loading 2608.25427v1…