Heading-Equivariant Lag Features for Mapping with Slow Chemical Sensors

Many chemical, gas and biological sensors respond slowly. A metal-oxide gas sensor can take a minute to recover, and deep-sea methane instruments lag by 12 to 35 minutes. When such a sensor rides on a moving robot, time lag turns into spatial error. The robot reports a plume peak well past where the plume actually is, and the error points along the direction of travel. Current practice copes in one of two ways. It drives slowly in sweeping spirals so the errors average out, or it accepts kilometre-scale mislocalization and uses statistics after the fact to say roughly where something happened.

This proposal starts from a simple observation that turns that error into a signal. Lag displaces every reading by roughly the platform's velocity times the sensor's time constant, so passes over the same patch of water or air from different headings disagree in a predictable, direction-dependent way. A learned map that stores each detected feature as a vector, with one part that stays fixed under changes of heading (where the source really is) and one part that rotates with heading and scales with speed (how far lag has pushed it), can therefore estimate the sensor's unknown time constant online and undo the smear. No laboratory calibration is needed. The robot also no longer has to creep along in spirals.

Core Claim

For a sensor with first-order response, the reading at time $t$ is an exponentially weighted average of the true concentration along the robot's recent path. Under roughly constant velocity, the centroid of that weighting sits a distance $v\tau$ behind the robot, which follows from the mean of an exponential kernel being $\tau$. A feature whose true location is $x$ is therefore reported at

$$\tilde{x} = x + R(\theta)\,(s\tau,\,0),$$

where $\theta$ is the heading, $s$ the speed, $\tau$ the sensor's effective time constant, and $R(\theta)$ the planar rotation by $\theta$.

Two consequences follow. The offset is equivariant to a known group: rotating the heading rotates the offset, and scaling the speed scales it. And $\tau$ becomes identifiable from any two non-parallel passes over the same structure, without knowing the binding rates.

We claim that an encoder trained to respect this structure will do two things that neither scalar deconvolution nor unconstrained learned world models do. It will estimate the field and $\tau$ jointly when the kinetic parameters are wrong or drifting. It will also extrapolate to survey speeds and headings outside its training range. That extrapolation is what lets a robot survey fast instead of in slow spirals.

Where This Sits

The problem is documented by work that motivates this proposal. Preston et al. measured 12-minute (SAGE) and roughly 35-minute (Pythia) lags on the vehicle Sentry and saw hydrothermal signals register 1.5–3 km laterally from the source. Their response was streaming change-point detection and cross-correlation with turbidity, which flags events but never inverts the lag. Lilienthal et al. handled a sensor that takes up to 70 s to recover by forcing a spiral sweep and Gaussian kernel smoothing. Their sweep is a hand-designed way of averaging over headings. We make the heading dependence explicit and exploit it rather than average it out.

Two works supply methods we reuse. DUSK inverts reversible binding kinetics with an ordinary differential equation coupled to an untrained convolutional prior. It raises recovered signal-to-noise from 1.31 dB under an equilibrium assumption to 12.52 dB. However, it runs offline on fixed microscopy frames, and it needs known forward and backward rates, which are exactly what drift in the field. We borrow its kinetic forward model to simulate sensors, but we do not assume its rates are known.

LASER moves sparse sensors using a latent world model and a reinforcement-learning policy. On 2D Navier–Stokes with 64 sensors on a 64×64 grid it reaches 0.434 reconstruction MSE against 20.27 for the fixed-layout AROMA baseline. Its sensors respond instantly, so lag-aware mapping is the piece it leaves out.

The representation comes from capsule networks. Sabour et al. introduced vector "capsules", whose length encodes whether an entity is present and whose direction encodes its pose. EquiCaps trains such vectors to be invariant in one part and equivariant in another under known transformations, reaching rotation $R^2$ of 0.78 on 3DIEBench. EquiCaps notes that it depends on known group actions. In our setting this limitation does not bite, because heading rotation and speed scaling are known exactly from navigation.

Data-quality work constrains the design. EngiAgent and the PDEBench findings on conservation error (Takamoto et al.) warn that neither bigger models nor reasoning loops recover physical parameters that are absent from the input. We take that seriously: our claim is only that $\tau$ is present in the data, encoded in heading-dependent disagreement.

Research Program

We will build a lag-aware mapping model with three parts.

The first is a sequence encoder over raw sensor readings, robot pose, and velocity. It emits, per spatial cell, a vector feature. The length of the vector is plume presence, an invariant block holds the corrected location and intensity, and an equivariant block holds the apparent offset.

The second part is training. The model learns on simulated fields passed through DUSK-style first- and second-order binding kinetics, with $\tau$ and rates randomized. Training uses three losses, in the spirit of EquiCaps. An invariance loss across passes over the same field at different headings and speeds ties the source estimate to a single answer. An equivariance loss requires the offset block to transform as $R(\theta)\,s$. A reconstruction loss re-applies the estimated kinetics to the corrected field and must reproduce the raw readings.

The third part is decoding. A readout pools the offset block across passes to produce a running estimate of $\tau$ with uncertainty, and outputs the corrected field.

We will test routing-by-agreement only where it plausibly earns its cost. That case is two nearby sources whose smeared tails overlap, where "explaining away" should assign each tail to one source; the open question raised in the research file's Gap 4 is whether slow decay tails instead keep false-plume capsules active. If routing gives nothing there, we drop it and keep plain equivariant vectors.

A secondary question concerns planning. For instruments with minute-scale lag, the robot chooses its next survey leg on a timescale where seconds of language-model deliberation cost nothing. Si et al. show reasoning length can inflate from 169 to 1,531 tokens, but even that fits inside a 12-minute recovery window. We will test whether a chain-of-thought planner improves leg selection. Chain-of-thought here means a language model writing out intermediate reasoning steps before answering, and the planner would follow a fixed protocol modelled on SoS1's structured steps: check offset consistency, update $\tau$, find unresolved sources, propose a heading. We treat this as a hypothesis with a kill criterion, not as part of the core contribution.

Baselines

The naive reference is to use raw lagged readings with Kernel-GDM-style Gaussian smoothing on straight survey lines, as in Lilienthal et al.. This shows how much smear there is to remove. We also run its spiral-sweep variant at matched mission time, because the spiral is the field's current answer to lag.

The strongest classical competitor is an extended Kalman filter with $\tau$ added to its state, so that the time constant is estimated jointly with a parameterized plume model from the same trajectories. It uses the same simulated kinetics and gets the same knowledge of pose and velocity as our model. Its plume prior is tuned on the training split with a search budget equal to ours. The nominal-$\tau$ version, given a deliberately wrong time constant, shows how fragile known-parameter inversion is. This mirrors the gap between equilibrium inversion and DUSK.

The strongest learned competitor is a LASER-style latent world model trained directly on lagged observations, with $\tau$ randomized over the same range. It gets the same architecture budget, training data, and GPU hours. Lagged sensing has no published result; LASER's 0.434 MSE on zero-latency 2D Navier–Stokes is the only reported anchor, and we will re-run LASER at zero lag to confirm our setup reproduces it before adding lag.

The key ablation is our exact network with the equivariance loss removed and the offset block replaced by ordinary scalar channels, with heading and speed supplied as inputs. This isolates whether the imposed structure, rather than capacity or data, does the work.

Our method beats these baselines if, at survey speeds 2–3× beyond the training range, it reduces source localization error below 10% of $s\tau$ while the learned baselines exceed it. It must also recover $\tau$ within 15% from four or fewer crossings. And it must match or beat the augmented-state Kalman filter in turbulent fields where that filter's plume prior is wrong.

Why not just…?

Why not just run a Kalman filter with $\tau$ in the state? This is the most serious alternative. It uses the same identifiability argument and could match us in laminar, single-source fields. Its weak point is that it needs an explicit field model, and Lilienthal-style Gaussian dispersion degrades under advection and turbulence. We will report the comparison separately for diffusion-dominated PDEBench fields and turbulent Navier–Stokes fields. If the filter ties us in the turbulent fields too, the honest conclusion is that the learned representation adds little, and we will say so.

Why not just train a world model on lagged data and let its memory absorb the lag? It likely works within the training range of speeds and headings, which is why it is a baseline rather than a strawman. The equivariance constraint matters only where the network must extrapolate: unseen speeds, unseen headings, and time constants outside the randomized range. The out-of-range speed sweep settles this directly. If the unconstrained model generalizes as well, the vector structure is unnecessary.

Why not just plan the next leg by expected information gain instead of chain-of-thought? An information-gain planner over our $\tau$ posterior is cheap and principled, and we expect it to be hard to beat. Evidence that structured reasoning helps comes from symbolic tasks, such as EngiAgent's feasibility rising from 7.55% to 64.15% on GPT-4o. That evidence does not transfer automatically to geometric decisions. We will run both planners on identical missions. If chain-of-thought does not reduce the number of crossings needed to localize all sources, we drop it.

Proposed Experiments

For synthetic fields we use PDEBench advection, diffusion-reaction, and 2D Navier–Stokes. We also use LASER's 64×64 Navier–Stokes setting, with sensor readings passed through first-order lag and DUSK's binding equation. Time constants are scaled so the ratio of $s\tau$ to plume size spans the MOX corridor regime (about 70 s) up to the Guaymas regime (about 12–35 min).

The metrics are corrected-field MSE, PDEBench conservation error (cRMSE), source localization error normalized by $s\tau$, $\tau$ estimation error, and mission time to localization.

For real data, we will run the RR2107 Guaymas transect, where the 0.8 methane–turbidity correlation gives a fast-responding reference, if the data can be obtained. No numeric localization result for that transect is reported on Lacuna. Lilienthal's 20 m corridor has no reported numeric error metric found on Lacuna, so it can serve only for qualitative checks.

There are three decisive checks. First, with a nominal $\tau$ off by 2×, does our $\tau$ estimate converge while the nominal-$\tau$ filter's localization diverges? Second, does the equivariant model beat its non-equivariant ablation at out-of-range speeds? Third, can a fast straight-line survey with crossing legs match spiral-sweep localization in less mission time? The third is the operational payoff. If it fails, the speed-versus-accuracy trade-off documented in this literature remains unbroken.