IROS 2026 Best Paper in the Field | Visual SLAM Frequently Loses Track Underwater, This Team Uses Robotic Arm Suction Cups for Localization

“

Under the murky seabed, robots can determine their position by “touching.”

Underwater sampling faces a very practical challenge:

Robots often don't know where they are.

There is no GPS underwater, and autonomous underwater vehicles (AUVs) mainly rely on inertial measurement units (IMUs) and Doppler velocity loggers (DVLs) for dead reckoning, which leads to error accumulation over time. Visual SLAM commonly used on land is not effective underwater due to light scattering, murky water, and sparse seabed textures, all of which significantly reduce the effectiveness of visual positioning.

A research team from the University of Heriot-Watt in the UK and the Japan Agency for Marine-Earth Science and Technology proposed a solution:

Using the physical contact between the robotic arm's suction cup and underwater objects as a form of positioning information.

This paper, titled “Contact-Aided Factor-Graph Localization for Underwater Sampling,” won

the Best Paper Award in the Safety, Security, and Rescue Robotics Track at IROS 2026.

01.

Why is visual localization prone to failure in low-altitude close-to-seabed operations

When performing underwater sampling, the AUV flies close to the seabed at low altitude, with the camera facing directly downward. In this scenario, the image is mostly flat and has sparse textures.

The study analyzed several issues encountered by monocular visual odometry in this configuration:

Translation parallel to the seabed is difficult to observe, scale is uncertain without reliable height information, and feature tracking is unstable. Starfish, a type of benthic organism, are small in size and sparsely distributed, often appearing only briefly in the camera's view, leading to large distance estimation noise and unreliable recognition.

The team conducted quantitative statistical analysis on the experimental data. In the port sequence, an average of 418 key points were detected per frame, with the minimum being 90 per frame, and only 42% of the key points were retained between frames. In the first water tank, although an average of 781.8 key points were detected per frame, the cross-frame retention rate was only 47%.

In the second water tank, a total of 4078 frames were analyzed. The data showed that the median value of optical flow amplitude was only 1.24 pixels, the disparity is insufficient to stably triangulate depth.

Under these conditions, COLMAP was unable to complete reliable sparse reconstruction initialization, **ORB-SLAM3, VINS-Mono, and ROVIO also repeatedly lost track or diverged.** The paper concludes that classical monocular SfM and visual SLAM methods are inherently ill-posed in this scenario.

02.

A single touch of the suction cup is equivalent to performing a loop closure

Since they couldn't see clearly, the team let the robot go 'feel'.

The idea is that when the suction cup grabs an object on the seabed, the position of the end effector can be precisely calculated by the forward kinematics of the robotic arm. This position has practical measurement significance and does not change with the observation angle. The team modeled this contact as a high-confidence constraint in a factor graph.

high-confidence constraint.

If there is already a previously built landmark nearby, such as a starfish detected visually, the system adds a bearing-distance factor between the current pose and this landmark. If there is no landmark nearby, it creates a new landmark based on the contact point. Both scenarios use a very tight noise model to ensure geometric consistency.

In this way, when the suction cup touches an object that has been previously recorded, it forms an implicit loop closure, and the entire process does not rely on appearance-based place recognition.

The entire framework is implemented based on GTSAM, with the backend using iSAM2 incremental optimization, integrating IMU, DVL, pressure sensor, magnetometer, downward-facing monocular camera, target detection, and robotic arm contact information.

The visual part has a key design: the system adaptively adjusts the covariance of the visual odometry based on the inlier rate. The inlier rate is first processed with an exponential moving average and then constrained within a certain range. This ensures that frames with poor visual quality do not bias the estimator.

Target detection uses YOLOv8, trained on a public starfish dataset. It estimates distance by combining the height of the detection box with a pinhole model. When the confidence is low or the detection box is too small, the measurement uncertainty is amplified.

In addition, the system uses

hybrid sliding window local bundle adjustment:

When there are visual landmarks in the window, the system optimizes along with the landmarks; when there are no landmarks, it only optimizes the pose. The system can also complete online initialization during motion without needing to stop first.

03.

Field tests in three types of scenarios: water pool, port, and simulation

The experimental platform is equipped with a JIMS-80 MEMS IMU, Wayfinder DVL, Paroscientific pressure-temperature sensor, magnetometer, and downward-facing monocular camera. The experimental scenarios include two water pools (100m×7.8m×4.35m and 40m×4m×2m), a port area of approximately 10m×10m×5m, and the Stonefish simulation environment.

In the first water pool, the team designed five trajectories. The ATE for straight-line motion is 0.0990 meters. After adding two 180° turns, the ATE increased to 0.2917 meters, indicating that planar rotation aggravates drift. After adding depth variation, the ATE dropped to 0.1954 meters.

When depth variation and starfish landmarks are both present, the yaw RMSE reaches its lowest at 0.0412 radians.

The ablation experiments show a clear contrast. On the straight trajectory, using only DVL, IMU, and pressure sensor, the ATE is 0.6780 meters, and the yaw RMSE is as high as 2.2396 radians; for the complete system, they are 0.0990 meters and 0.0524 radians, respectively. Removing the camera results in an ATE of 0.2798 meters, and removing local bundle adjustment results in an ATE of 0.2270 meters.

The Stonefish simulation provides ground truth. Trajectories with sufficient excitation, such as lateral motion and yaw variation in depth, have ATEs below 0.1 meters. The ATE is largest for continuous planar rotation, at 0.998 meters, but the relative pose error is only 0.0032 meters. In the trajectory with a half-eight shape and two dives to starfish, after combining landmarks and suction cup contact, the starfish revisit error is only 0.06 meters.

The second water pool focuses on testing the effect of contact. The robot first detects the starfish, then approaches, and finally activates the suction pump for contact.

The most telling is the revisit error:

Without landmarks, it is 0.58 meters; after adding visual landmarks, it drops to 0.50 meters, and further decreases to 0.45 meters with the addition of suction pump contact.

The port environment is closer to real sea conditions, with turbid water and changing lighting. The system's ATE here is 0.8223 meters, which is larger than in the pools, but the trend of revisit error change is consistent:

from 1.23 meters to 1.12 meters, then to 0.90 meters.

The team concluded that in perception-limited underwater tasks, drift is largely an observability issue. Planar rotation amplifies drift, while vertical motion improves the observability of scale and heading. Visual landmarks mainly help stabilize orientation, and contact constraints can consistently reduce revisit errors near the operational position.

Next, the team plans to combine this contact-assisted localization with some high-frequency voxel mapping methods for dense mapping and operation planning in benthic work areas.

END

Industrial robot companies

Eston Automation

|

Efte Robot

|

Fao Robot

|

Yuexiang Robot

|

Jieka Robot

|

Songling Robot

|

Luoshi Robotics

|

A Tongmu Robot

|

JiZhiJia

|

Hikvision Robot

|

Yifei Technology

|

Ailite Robot

Service and special robot companies

Yijiahe

|

Jingpin Tieshang

|

Qitem Robotics

|

Shihe Robot

|

Pudu Robot

|

Shiluo Robot

|

Kuma Technology MAMMOTION

Humanoid robot companies

UBTech Robotics

|

Unitree

|

Yunchedu

|

Xingdongjinyuan

|

Weijing Robot

|

Zhujidongli

|

Leju Robot

|

Daxiang Robot

|

Magic Atom

|

Zhongqing Robot

|

Pasi Ni Perception

|

Cyborg Robot

|

Digital Huaxia

|

Fourier Intelligent

|

Tianlian Robot

|

Kepler Humanoid Robot

|

Lingbao CASBOT

|

Qingbao Robot

|

Zhejiang Humanoid Robot Innovation Center

|

Dongyi Technology

|

Zhishi Technology

|

PNDbotics

|

Zhuoyide Robot

|

Qinglang Intelligent

|

Galileo GALILEO

|

Songyan Power

|

Tianji Intelligent

|

Woaan Robot

|

LiGong Huahui

|

Jiace Jinhua

Embodied intelligence companies

Cross Dimension Intelligent

|

Galaxy Universal

|

Qianxun Intelligent

|

Lingxin Qiaoshou

|

Ruierman Intelligent

|

Microintelligence

|

Tuisheng Technology

|

Zhongke Silicium Records

|

Shutu Technology

|

Lingqiao Intelligent

|

Xingchen Intelligent

|

Qiongche Intelligent

|

Fangzhou Infinite

| iFLYTEK |

Beijing Humanoid Robot Innovation Center

|

National and Local Co-built Humanoid Robot Innovation Center

|

Daimeng Robot

|

Shibitu Robot

|

Xinghaitu

|

Yuequan Bionics

|

Zeroth Power Robot

|

Zhongke ShenGu

|

Zhi Fangping

|

Daka Robot

|

Haocun Technology

|

Ju Shi Intelligent

|

Xynova Xiner Future

|

Feixi Technology

|

Weilu Power

|

Boden Intelligent

|

Qianjue Technology

|

Lingsheng Technology

|

Ji Cui Intelligent Manufacturing

|

Xinbeit Technology

|

Chen Hunxian Technology

|

Dexmal Original Power

|

Youliqi

|

Self-variable

|

Ruiyan Zhikong Lingqiao Hand

|

Qiwu Technology

|

RoboScience Machine Science

|

Zhongke Fifth Dynasty

|

Critical Point

|

Danghong Technology

|

Qiaojie Shu Wu

|

Vbot Veta Power

|

Tashan Technology

|

Ju Brain Pan Shi

|

Youai Zhijhe Robot

|

Zhixing Jian

|

Aimiao Robot

|

Lu Ming Robot

|

Quantum Intelligent

|

Ruoyu Technology

|

Tiehu Robot

|

Zhi Wuji

Medical robot companies

Yuanhua Intelligent

|

Tianzhihang

|

Sizhe Rui Intelligent Medical

|

Jingfeng Medical

|

Tuo Dao Medical

|

Zhenyida

|

Shuru® Robot

|

Luosenbot

|

Shuimu Dongfang

｜

Kangnuo Sistem

|

Disight Medical

Upstream industry chain companies

Lvde Harmonic

|

Yinshi Robotics

|

Kunwei Technology

|

Muta Intelligent

|

Qingpou Vision

|

Benmo Technology

|

Lan Dian Touch Control

| Xinjingcheng Sensor

|

BrainCo Strong Brain Technology

|

Yuli Instrument

|

Jiya Jingji

|

Sila Technology

|

Shenyuan Sheng

|

Feipu Navigation Technology

|

Inks

|

Juxi Intelligent Drive

|

Lingyun Guang Yuanke Vision

|

Xuanji Power

|

Yiyou Technology

| Ruiyuan Precision |

Lingzhu Era

|

HIT Huawei Ke

|

Xinghui Sensor

|

Lingdi Technology

|

Quan Zhi Bo

|

CubeMars Robot Power

|

Wanglong Robot Elevator

|

Hangkai Microelectronics

|

Wutong Sensing and Control