When contact is a fix: unpacking contact-aided AUV localization

A localisation paper that treats touching the seabed as data

Michele Grimaldi, Yosaku Maeda, Hitoshi Kakami, Ignacio Carlucho, Yvan R. Petillot and Tomoya Inoue posted “Contact-Aided Factor-Graph Localization for Underwater Sampling” to arXiv on 27 August 2026. The paper addresses a specific and awkward operating point: an autonomous underwater vehicle flying low over the seafloor to sample it, where the down-looking camera stares at featureless, near-planar sediment. In that regime the authors identify three failure modes for visual odometry – scale ambiguity, lateral degeneracy, and inconsistent feature tracking – and note that inertial–DVL fusion, on its own, offers no mechanism to correct structural drift.

Their proposal is to treat the moment the vehicle’s suction-based manipulator makes physical contact with the seabed as an informative geometric constraint inside a smoothing-based estimator. The framework fuses those contact events with adaptive visual odometry, learned object detections and the standard onboard sensor suite. Visual relative-pose factors and landmark bearing-range factors are uncertainty-scaled by inlier statistics so weak frames cannot destabilise the solution, while contact events enter as high-confidence factors that induce implicit loop closures without appearance-based place recognition. The system initialises online during motion, and the authors report reduced trajectory drift and improved object-revisit accuracy across tank, harbour and simulation trials, measured against filtering-based navigation and contact-free graph formulations.

This is a research contribution, not a standard, and we should be plain about that from the outset. What makes it worth a survey manager’s attention is the underlying idea, because it names a real gap in how we navigate close to the bottom and proposes a mechanism that transfers cleanly to work-class ROV and intervention contexts.

What the method actually does, in plain terms

Start with the estimator. Most subsea navigation stacks in service today are filtering-based – an extended Kalman filter blending INS, DVL bottom-lock, depth and, where available, USBL or LBL aiding. A filter marginalises the past: once a state is folded in, its history is gone, and a correction that arrives late cannot be applied retrospectively to earlier poses. A factor graph does the opposite. It keeps the trajectory as a set of pose nodes connected by constraint edges – each measurement is a factor carrying its own covariance – and re-optimises the whole structure when new information arrives. That is what makes loop closure and delayed corrections tractable, and it is the reason the wider robotics community moved to smoothing formulations for anything involving revisits.

The first practical idea in the paper is adaptive weighting of the visual factors. Over a bare sand or mud plain, a monocular down-looker cannot resolve scale, and translation parallel to the seabed is weakly observable – the classic degeneracy where the image barely changes as the vehicle slides sideways. Rather than let a frame with a handful of tracked features drag the solution off course, the authors inflate that factor’s covariance in proportion to its inlier count. A strong frame informs the estimate; a weak one is admitted with wide uncertainty and does little harm. This is disciplined robustification, and it is the correct instinct for the environments where optical aiding is least trustworthy. We have argued before that buying position confidence from cameras in the absence of acoustics is only as good as the frame-by-frame quality gating behind it, a point we develop in our analysis of visual navigation as a substitute for acoustic aiding.

The second idea is the interesting one. When the suction manipulator contacts a physical point – a sampling target, a marker, a distinctive object detected on the seabed – the estimator gains a geometric constraint tying the vehicle to that point’s location. Return to the same object later and contact it again, and the graph has an implicit loop closure: the vehicle is known to be at the same physical place, so accumulated drift between the two visits can be squeezed out. The elegance is that this needs no visual place recognition. Appearance-based loop closure fails precisely where this method operates, because a featureless seabed produces no distinctive bag-of-words signature to match against. Contact sidesteps the perception problem by using the body of the vehicle as the sensor. It is the same principle exploited by zero-velocity updates in legged and pedestrian inertial navigation, transplanted to a manipulator tip on the bottom.

What it means for close-range sampling and sample georeferencing

The operational problem this speaks to is well known to anyone who has run a low-altitude AUV or a work-class ROV in survey mode near the bottom. DVL bottom-lock gives you velocity over ground, not position, and integrating velocity accumulates error. Heading error from the gyro couples straight into cross-track position growth, so the along-track and across-track error ellipse walks even with a healthy Doppler solution. Absolute correction has to come from somewhere – USBL or LBL aiding, a terrain fix, or a loop closure. In deep, featureless terrain none of the geometric aids is generous.

USBL is the usual answer, but at extreme proximity to the seabed the geometry works against you. Grazing incidence, refraction through stratified water and multipath off the bottom degrade the acoustic fix exactly when the vehicle is doing delicate close-range work, a set of constraints we have unpicked in our note on USBL behaviour in extreme-proximity positioning. A contact constraint offers a second, independent way to bound relative drift that does not depend on a clean acoustic path back to the surface or to seabed transponders.

Why should a survey manager care about a few centimetres of drift on a sampling run? Because the sample is only worth what its position is worth. A core, a grab or a biological specimen whose recovered location is uncertain to metres is scientifically and commercially weaker than one tied down to decimetres, and repeat sampling – returning to the same site across a campaign or across years – depends entirely on revisit accuracy. IHO S-44 frames total horizontal and vertical uncertainty for hydrographic survey; the same logic applies to sample georeferencing, where the deliverable is the position of a physical thing you have taken from the bottom. Object-revisit accuracy, the metric the paper reports, is the property that governs whether you can find that site again and whether your before-and-after comparison is honest.

There is a broader signal here for anyone specifying autonomous seabed intervention. The industry has spent two decades improving the sensors we point at the world. This work argues for treating physical interaction – something an intervention vehicle does anyway – as a localisation primitive in its own right. That reframing matters for scientific coring, benthic sampling, cable and pipeline touch-down work, and any task where the vehicle already makes deliberate, repeatable contact with a known feature.

Gaps and gotchas before anyone writes it into a spec

The honest engineering position is that this is tank-, harbour- and simulation-grade evidence. It is a credible demonstration, not a qualified navigation mode, and the distance between the two is where the real work sits.

Contact factors are high-confidence, which cuts both ways. A tight covariance on a correct contact constraint corrects drift efficiently. A tight covariance on a wrong one – a mis-associated contact, the manipulator touching a different object than the estimator believes – injects a confident error that the smoother will trust and propagate through the whole trajectory. Data association is the dominant failure mode. A soft, wrongly-weighted visual factor degrades gracefully; a hard, wrongly-associated contact factor does not. Any implementation needs a statistical gate – a chi-square consistency test on the contact factor against the prior – before the constraint is admitted, and a mechanism to demote or drop contacts that fail it.

Soft sediment breaks the static-point assumption. An implicit loop closure assumes the contacted point sits where it did on the previous visit. On cohesive mud, on drift-prone fine sand, or on an object that the suction event itself disturbs, that assumption weakens. The contact geometry carries its own uncertainty – manipulator compliance, seabed deformation, the object shifting under load – and none of that is captured by treating contact as a near-perfect fix.

Learned detection degrades in the same conditions as the camera. The object detections that seed contact targets come from the same optically-poor scene that defeats the visual odometry. False positives and missed detections do not disappear because the detector is a neural network; turbidity, low altitude and flat backgrounds stress it. The detection confidence threshold is a tuning parameter with direct navigation consequences, and it needs to be version-controlled alongside the model weights.

Loop closures bound relative drift, not absolute position. This is the point most likely to be missed on the deck. Contact-induced closures make the trajectory internally consistent, but a self-consistent graph can still float bodily relative to the geodetic frame. Without at least one absolute fix – a USBL or LBL update, a surveyed datum on the seabed – there is nothing tying the whole solution to WGS84 or to your project grid. Contact aiding sharpens the shape of the trajectory; it does not, on its own, place it. For a survey deliverable that must sit in a national or project datum, the anchoring fix is not optional.

Contacts are sparse. Between physical interactions the estimator is back on INS, DVL and whatever visual factors survive gating, so drift grows in the gaps and is only pulled back at the next contact. The method rewards mission profiles with frequent, deliberate revisits and helps least on long transit legs.

What to do with this now

Treat contact-aided localisation as an additional constraint to be qualified, not a navigation mode to be trusted out of the paper. Concretely:

• Keep the INS/DVL backbone and an absolute aiding source. Anchor the factor graph to the geodetic frame with at least one USBL or LBL fix, and re-anchor on a defined schedule for any run whose product must sit in a project datum. Do not let a self-consistent but unanchored graph masquerade as a georeferenced trajectory.
• Instrument the contact itself. Gate contact factors on force/torque or suction-pressure evidence that a genuine, stable contact occurred, and reject events that do not clear a defined contact-quality threshold. A visual detection alone is too weak to authorise a high-confidence factor.
• Put a consistency gate in front of every contact factor. Apply a chi-square test against the prior before admission, log the residual, and demote contacts that fail rather than forcing them into the solution.
• Log the graph, the covariances and every factor’s weight for post-mission QA. The audit trail is the integrity case. A survey party chief signing off sample positions needs to see which constraints shaped the trajectory and how confident each was – the same evidential discipline we apply to GNSS integrity in rigorous offshore positioning QA.
• Benchmark against a stated accuracy target. Set an object-revisit and position-uncertainty budget derived from the IHO S-44 order the campaign is contracted to, and prove the method meets it in representative sediment before it carries a deliverable.
• Run a qualification programme in the sediment you will actually work. Tank and harbour results do not transfer to soft deep-water mud without evidence. Characterise contact-point stability, detection performance in expected turbidity, and drift growth between contacts under representative currents.

The contribution earns its place because it names a genuine weakness in low-altitude subsea navigation and answers it with a mechanism the vehicle already has the hardware to exploit. The engineering caution is equally genuine: a high-confidence constraint is an asset when the association is right and a liability when it is wrong, and it corrects the shape of a trajectory without ever placing it in the world. Both truths belong in the specification.

---

Based on: Contact-Aided Factor-Graph Localization for Underwater Sampling