When a robot enters a damaged reactor, the computer model of the radiation field is often badly wrong. At Fukushima Daiichi Unit 1, a detailed Monte Carlo simulation predicted 1.4 Sv/h at the investigation robot's position, but the robot measured 7.2 Sv/h. The gap was closed afterwards by multiplying one class of sources by a single fitted factor of 5.14 (Okumura et al.). The robot followed a pre-scripted route, and the model was fixed only once the robot was back out. This proposal makes the robot choose where to measure, while it is still inside, so that each reading best corrects the physics model. The catch is that the robot's own accumulated radiation dose, which destroys its electronics, is computed from the same uncertain model it is trying to fix.
The key observation is that photon transport is linear in source strength. Once a Monte Carlo code has computed how much dose each source region delivers to each possible robot position, the dose rate anywhere is a weighted sum of unknown per-region source intensities. Under Gaussian beliefs, three quantities then have closed forms: the updated belief after each reading, the information a candidate measurement would provide, and the probability that a planned path exceeds its dose budget. That makes it cheap to plan several steps ahead with a chance constraint on dose, which existing active-calibration methods cannot do cheaply. We ask whether this physics-structured active calibration reaches a given accuracy on dose predictions, including predictions at places the robot never visits, at substantially lower robot dose than surveys, greedy information seekers, or model-free mapping.
Core Claim
Let $s \in \mathbb{R}^K$ be the unknown intensities of $K$ source regions (63 in the Unit 1 decomposition). Let $a(x) \in \mathbb{R}^K$ be the precomputed transport response at robot position $x$, meaning the dose rate each region delivers per unit intensity. A detector reading is then
$$y(x) = a(x)^\top s + \delta(x) + \varepsilon,$$
where $\delta(x)$ is a smooth discrepancy field. It is modeled as a Gaussian process, a flexible prior over functions, and it absorbs anything the region decomposition misses, such as unmodeled hot spots. The term $\varepsilon$ is counting noise whose variance grows with the rate.
Put a Gaussian prior on $s$ around the simulation's nominal inventory. The belief update is then an exact linear-Gaussian update. The information a reading at $x$ adds is $\tfrac12\log(1+v(x)/\sigma^2(x))$, where $v(x)$ is the current predictive variance and $\sigma^2(x)$ is the noise variance. The robot's cumulative dose along a path $\gamma$ is also a linear function of $s$ and $\delta$:
$$D(\gamma)=\int_\gamma \big(a(x)^\top s+\delta(x)\big)\,dt.$$
Its posterior mean $\mu_D$ and standard deviation $\sigma_D$ are therefore available in closed form. The planner maximizes the information gained along a multi-step path subject to $\mu_D(\gamma)+z_\alpha\sigma_D(\gamma)\le B - D_{\text{return}}$. Here $B$ is the hardware dose budget, $z_\alpha$ is the standard-normal quantile for the allowed violation probability $\alpha$, and $D_{\text{return}}$ is the conservatively estimated dose of getting back out.
This creates a coupling that no cited work models. Measurements that are informative about the dangerous regions also tighten $\sigma_D$ for paths near those regions. So learning can make aggressive paths admissible later in the mission, rather than information and safety simply trading off.
Relation to Prior Work
Three works motivate the gap without closing it. Okumura et al. supplies the forward-model structure we reuse: region decomposition, inventory and activation codes, and PHITS photon transport. Its calibration was post-hoc, used a single scalar, and followed a fixed path, and it left the reactor pressure vessel interior uncalibrated. Ricciardi et al. interlaces experiment choice with Bayesian calibration and reports posterior generalized variance "orders of magnitude" below static designs. However, it needs Gaussian process surrogates and Laplace approximations, it is stated to become prohibitive beyond about five parameters, and its "experiments" carry no cost that endangers the apparatus. We exploit transport linearity to handle about 63 parameters exactly, and we add a hazard budget. Schwager, Dames et al. show that robots maximizing mutual information naturally avoid lethal hazards. Their approach is myopic, uses flat priors with no physics model, and was evaluated only in grid simulation. We keep their insight that attrition destroys future information, but make the hazard a learned physical field that is also the calibration target.
Two works constrain the design. Wu et al.'s ID3QN plans for low dose on a known, static radiation map, and its authors note lengthy training and fragile hyperparameters (companion entry). A policy trained offline on a field that is off by 5x is the wrong tool when the field is being revised online, so we plan over the posterior instead. AIPPMS supplies the planning machinery we reuse: Monte Carlo tree search over a partially observed state, with pruning that guarantees a feasible return. Its budget, however, is deterministic battery consumption, not an uncertain exposure. Finally, Kaheman et al. show that sparse symbolic correction terms on top of physics beat both uncorrected physics and black boxes. Our $\delta(x)$ plays that residual role. We use a Gaussian process rather than a sparse symbolic library because the paper itself flags spatially non-local mismatches as a failure mode of sparse libraries.
Research Program
The work has three parts. First, a response-matrix pipeline. For a plant geometry, run Monte Carlo transport once per source region onto a candidate grid of robot positions to obtain $a(x)$. This is an offline cost, identical in kind to what Okumura et al. already computed. Second, the online estimator. It performs exact Gaussian updates on $(s,\delta)$ with rate-dependent noise, and enforces non-negativity of $s$ by a log-transform variant as an ablation. Third, the planner. It is a receding-horizon tree search over waypoints. Information is scored in closed form, the dose chance constraint is checked at every node, and branches without a safe return are pruned.
We also study misspecification directly. We inject ground-truth sources outside the region decomposition, and we inject detector saturation. We then measure whether $\delta$ absorbs these errors or whether the source estimates get corrupted, because a biased $s$ makes the dose constraint itself unsafe.
Baselines
The naive reference is the nominal simulation with no calibration. It tells us how bad planning on the uncorrected model is, measured by prediction error and by true dose incurred when the planner trusts the nominal map. The historical practice is a pre-planned lawnmower survey at the same dose budget, followed by offline least-squares fitting of a single global multiplier, reproducing the 5.14-style correction of Okumura et al.. A stronger variant of this baseline fits all $K$ region intensities offline from the same survey data. Surveys are truncated at the same budget $B$.
The strongest active competitor is a greedy, one-step information seeker with a dose penalty, in the spirit of Schwager, Dames et al.. It uses the same physics posterior, so the comparison isolates planning horizon and the chance constraint rather than the model.
A model-free competitor uses a Gaussian process dose map with informative path planning, with the same planner, budget, and kernel-tuning budget. It isolates the value of the physics prior.
For dose-aware navigation, ID3QN (Wu et al.) is trained on the nominal map and run on the true map. Wu et al. report a 23.5% dose reduction versus DQN, and algorithm scores of 5,025 versus 7 in simple grids and 4,413 versus 3 in complex grids. We reimplement ID3QN on their grid setting to confirm we reach comparable behavior before using it.
For active calibration in a radiation field, we found no reported result on Lacuna. The closest numbers are from Ricciardi et al., which reports only a qualitative orders-of-magnitude variance reduction on material tests, and AIPPMS, which reports an average reward of 555.3 for POMCP versus 295.7 for the naive baseline on search-and-rescue. Neither is directly comparable to our task.
The ablations are as follows. We drop the chance constraint and use the mean dose only. We drop $\delta$. We collapse $s$ to a single scalar. We replace the exact linear update with a Gaussian process surrogate plus Laplace approximation, as in the interlaced calibration method, to measure what linearity buys in decision latency.
All methods share the same response matrices, sensor noise model, candidate waypoints, dose budgets, and a fixed hyperparameter-tuning budget on held-out worlds. Beating a baseline means lower held-out log-dose prediction error at equal budget, with violation rate no higher than the target $\alpha$. We will state in advance that the proposal must reach the lawnmower's final error at no more than half its robot dose to count as a meaningful win.
Why not just…?
Why not just run a survey and calibrate offline? When the budget is generous and the site is open, a dense survey will eventually pin down all 63 intensities. That is a real threat, so we sweep $B$ from tight to loose and report where the curves cross. The evidence against surveys under tight budgets comes from two places. The B1 route left the pressure vessel interior unresolved (Okumura et al.). Static designs also lost by orders of magnitude in posterior variance in Ricciardi et al.. If the crossover budget turns out to be small, we will report active calibration as a niche tool rather than argue around it.
Why not just map the dose field with a Gaussian process and skip the transport model? A black-box map can interpolate between visited points, but it has no basis for predicting dose behind shielding or inside regions the robot must never enter. Those predictions are what decommissioning planners and later robots need. The held-out test includes exactly such unvisited locations. If the model-free map matches us there, the physics prior is not earning its offline cost. Kaheman et al. found physics-plus-residual models outperform both extremes, which suggests it will earn it, but that finding came from a mechanical system, not a radiation field.
Why not just be greedy with a dose penalty? Greedy information seeking is simple and, by Schwager, Dames et al., already hazard-averse. The weakness is that a one-step view cannot plan measurements that first shrink dose uncertainty and thereby make an informative inner region safely reachable, and it has no explicit return guarantee. The comparison with the same physics posterior isolates exactly this. If the gap is negligible, the chance-constrained planner is unnecessary complexity.
Why not just recalibrate the map and hand it to ID3QN? That works for traversal after calibration is done, and we use it that way. But ID3QN needs retraining whenever the map changes (Wu et al.), and it optimizes low dose, not information, so it cannot drive the calibration itself.
Proposed Experiments
In simulation, we build PCV-like geometries with PHITS-style region decomposition. Ground-truth worlds are drawn by perturbing region intensities log-normally, with spreads that include the observed 5x underprediction. Some worlds also get off-decomposition hot spots, with sources ranging up to the several-hundred Sv/h debris levels reported by Okumura et al.. The metrics are held-out log-dose error split into visited-neighborhood and unreachable locations, calibration coverage of the 90% predictive intervals, true cumulative robot dose, empirical budget-violation rate, and planning latency. We also run the Wu et al. grid environments to confirm the planner does not lose to ID3QN on pure low-dose traversal when the map is correct.
For physical validation, a small wheeled robot carrying a calibrated dosimeter operates in a shielded laboratory with sealed low-activity Cs-137 check sources. Absorbed dose is scaled so that a notional budget stands in for hardware lifetime. Source strengths are deliberately mis-specified to the planner, and some sources are placed outside the modeled regions.
Three outcomes are decisive. The first is whether the proposed method reaches survey-level accuracy at half the dose. The second is whether its violation rate stays at or below $\alpha$ under injected misspecification, and whether dropping $\delta$ breaks this. The third is whether per-region calibration beats the single scalar multiplier at unreachable locations. A failure on the second check would be the most important finding, because it would mean that uncertainty-aware safety guarantees inherited from a wrong physics model cannot be trusted.
References
- A method for the prediction of the dose rate distribution in a primary containment vessel of the Fukushima Daiichi Nuclear Power Station
- Bayesian Optimal Experimental Design for Constitutive Model Calibration
- A Multi-robot Control Policy for Information Gathering in the Presence of Unknown Hazards
- A Novel Path Planning Approach for Mobile Robot in Radioactive Environment Based on Improved Deep Q Network Algorithm
- Adaptive Informative Path Planning with Multimodal Sensing
- Learning Discrepancy Models From Experimental Data