When several robots map an unknown building together without a central server, they can only exchange small amounts of data when they come within radio range of one another. Today's peer-to-peer mapping systems decide what to send using a fixed geometric rule, such as "send the loop closures that most strengthen the shared map." A loop closure is a match showing that two robots saw the same place. The fixed rule ignores two things. First, the world contains people and moving objects, and tracking them competes with map building for the same scarce bandwidth. Second, the human supervising the team cares about different things at different moments, and those concerns combine in non-additive ways. Faster coverage is worthless if the robots stop reporting where people are, and a newly spotted hazard can suddenly make one corridor the only thing that matters.
We propose to treat each rendezvous as a small decision problem: choose which packets to transmit under a byte budget so as to maximize a learned model of the operator's utility. That model has two properties. It is non-linear, so it can represent thresholds and interactions between objectives. It also separates a persistent core, such as safety, from a transient focus that is allowed to drift. Because the robots make this choice a few times per minute rather than at camera rate, a slow but expressive Bayesian model becomes affordable. The question we test is whether this preference-driven allocation delivers more operator utility per byte than geometry-only prioritization, linear weights an operator can adjust, and uniform time-discounting. We also test the cheaper alternatives head-on, because they might be good enough.
Core Claim
Under tight peer-to-peer bandwidth in dynamic scenes, how the link is split between three kinds of content should depend on a posterior over operator utility. The three kinds are geometric loop closures, tracks of moving objects, and semantic submaps (map pieces labeled by object type). The operator utility in that posterior is both non-additive across objectives and split into a persistent part and a drifting part. We claim two specific advantages, and each can be falsified. The first is that a non-linear utility model beats linear scalarization when the operator's true trade-offs contain thresholds, and ties it when they do not. The second is that splitting utility into a persistent core and a recent focus beats exponential forgetting when drift coexists with a constraint that never expires.
The work builds on several pieces of prior work and departs from each. Swarm-SLAM ranks candidate inter-robot loop closures by how much they increase the algebraic connectivity of the pose graph, the Fiedler value $\lambda_2$, which measures how well-knit the shared map is. That ranking is our geometric signal, but in Swarm-SLAM it is the entire objective. It is fixed in advance and treats moving content as outliers. Kimera-Multi provides robust distributed optimization that rejects false matches and explicitly leaves dynamic objects to future work. Dynamic-scene systems show that modeling motion beats masking or ignoring it, but they are single-robot and heavy. De4D-SLAM runs at 0.47 FPS, and Dream-SLAM needs diffusion inpainting on a large GPU. On the preference side, Promptable Behaviors personalizes a five-objective navigation policy but assumes a linear, episode-static weighting and names linearity as its limitation. Non-Stationary DPO handles drift with exponential discounting of old feedback, which cannot tell an obsolete wish from a lasting constraint. None of these decides what a robot team should transmit based on what a human currently wants.
Research Program
The system has three layers. The bottom layer is an off-the-shelf decentralized mapping backend, Swarm-SLAM. We pair it with a cheap dynamic front end in the spirit of point-correlation DSLAM, which separates static from moving points by checking whether pairwise point distances stay consistent and reports about 2 ms per frame. We deliberately do not use a heavy 4D Gaussian representation, because the contribution is about allocation, not reconstruction. Each robot keeps lightweight tracks of moving objects, such as people and carts, alongside its pose graph.
The middle layer is the allocator. At each rendezvous a robot holds a pool of candidate packets: loop-closure candidates scored by their $\lambda_2$ gain, dynamic-object track updates, and semantic submap deltas. Each packet has a byte cost. The allocator predicts a mission-state summary vector $z(S)$ that would result from sending packet set $S$. That vector includes connectivity gain, per-zone coverage, the age of the freshest report on each tracked object class, and proximity of moving objects to declared hazard zones. The allocator then solves a knapsack problem under budget $B$ against one posterior sample of operator utility, which is Thompson sampling: drawing a plausible utility from the posterior and acting optimally for that draw.
The top layer is the utility model, and it carries the new modeling content. We write
$$U_t(z) = C(z) + F_t(z),$$
Here $C$ is a persistent core fit to all feedback without forgetting, and $F_t$ is a focus term fit only to feedback since the most recently detected change point. Both are sums of regression trees with a Bayesian prior, following BFTS, whose additive regression trees capture thresholds and interactions without assuming linearity and come with a regret bound that does not require it.
Feedback is occasional and cheap for the operator. Every few minutes the operator ranks a slate of about five rendered map-state summaries. The likelihood is a ranked-choice model from Beyond Pairwise rather than pairwise Bradley-Terry, motivated by Promptable Behaviors' finding that picking from a group of five is 17.8% more effective than pairwise choice. A comparison is assigned to the core only if it agrees with core predictions across change points. This is how a lasting safety constraint survives a sudden shift of focus. The latency objection that BFTS raises for control loops does not apply here, because MCMC runs between rendezvous, on the order of seconds.
The operator talks to one robot, so the preference posterior must also spread peer to peer. Robots exchange a handful of posterior tree-ensemble samples, which are small. The bytes spent on those samples are charged against $B$ like any other packet, so the method pays for its own coordination.
Baselines
The do-nothing reference is Swarm-SLAM's spectral prioritization as published. Moving content is rejected as outliers, and whatever budget remains goes to semantic submaps in arrival order. It shows what current decentralized mapping gives an operator without any preference model. The strongest existing component for the geometric side is the same spectral ranking with dynamic tracks added under a fixed hand-tuned share of bandwidth. It is the most competitive geometry-first design because it inherits Swarm-SLAM's map-accuracy behavior, which reports under 100 MB of total communication for a 475-meter, three-robot real mission. For bandwidth framing, Kimera-Multi reports up to 70% less communication than centralized baselines. We run both on identical simulated missions, link traces, and byte budgets. The fixed share is tuned by grid search with the same number of evaluation episodes the learned method uses for its own hyperparameters.
The strongest preference baseline is linear scalarization in the style of Promptable Behaviors. Utility is a weighted sum of the same summary features, weights are inferred from the same ranked slates, and the same allocator and Thompson sampling are used. Promptable Behaviors reports 93.5% cosine similarity to expert weights in AI2-THOR, but no reported result was found on Lacuna for preference-driven allocation in multi-robot mapping. We add an oracle-slider variant in which the simulated operator directly resets true linear weights whenever its intent changes. For drift, the baseline is the non-linear model with NS-DPO-style exponential discounting in place of the core/focus split, with the discount tuned on held-out drift schedules. NS-DPO reports a length-controlled win rate of 8.93 against DPO's 7.29 on UltraFeedback-LM, and reports that stationary DPO falls to roughly 50-55% reward accuracy after a preference flip. Those numbers come from text alignment and serve only as motivation.
Two ablations isolate the key ingredients. One keeps the core/focus split but replaces the trees with linear terms, which tests non-linearity. The other keeps the trees and removes the split, which tests the drift structure. Beating a baseline means higher time-averaged true operator utility, reported as regret against an oracle allocator that knows the true utility, at equal bytes and equal operator queries. The difference must hold with confidence intervals over at least 30 seeded missions per condition. Static-map trajectory error must not degrade by more than a pre-registered 10%.
Why not just…?
Why not just let the operator set linear weights with sliders? This is the most serious threat. If the operator's real trade-offs are close to additive, sliders are transparent and need no learning. The non-linear literature argues they are not: Zhou et al. show that gains in one objective cannot buy back an intolerable loss in another, and Lin et al. show that the value of one attribute depends on the others. Neither result is about robots, though. The experiment therefore includes a control family of truly linear operators, where we expect sliders to tie or win. We also include threshold and interaction operators, where they should fail because no fixed weight vector can express "coverage counts only while person-report latency stays under ten seconds." If sliders match us on the threshold operators too, the non-linear model is not worth its cost, and we will report that.
Why not just send everything and skip prioritization? At Swarm-SLAM's reported volumes, bandwidth might seem plentiful. The target regime is degraded links of the kind seen in subterranean deployments such as LAMP. We sweep the budget $B$ from unconstrained down to a few kilobytes per rendezvous and report the budget at which allocation stops mattering. If that crossover sits below realistic link capacities, the problem is less pressing than we claim.
Why not just discount old feedback exponentially? It is simpler and has worked for language models. The NS-DPO authors' own analysis, however, requires knowing the total amount of drift to set the discount, and uniform forgetting erodes a persistent safety constraint along with the stale focus. The discount ablation, run on drift schedules that include a never-expiring constraint, settles this. If tuned discounting matches the split model, we drop the split.
Why not just ask an LLM agent to interpret the operator's instructions? AcCoRD reports that even the strongest model satisfies every preference less than 30% of the time, and scores fall as low as 0.09 on triggered preferences, meaning desires that appear only once the agent exposes a particular state. That is exactly the "operator sees a hazard and refocuses" case. We do not run this baseline at scale, which is a stated limitation.
Proposed Experiments
The main testbed is multi-robot dynamic indoor missions generated with GRADE in Isaac Sim, which animates humans and clutter. Each mission uses three to five robots with emulated range-limited links and per-rendezvous byte budgets. Simulated operators are drawn from three utility families: linear, threshold/lexicographic with safety first, and pairwise-interaction. Drift schedules follow AcCoRD's taxonomy, adapted to space: sudden refocus, gradual shift, and triggered preferences that activate when a hazard first enters a shared map. The primary metric is cumulative utility regret against the oracle allocator. Secondary metrics are static-map trajectory error, tracking error and staleness for objects the operator cares about, total bytes including the cost of spreading the posterior, and operator queries per hour.
Two checks ground the components outside simulation. On TUM RGB-D and Bonn walking sequences, we confirm that the cheap dynamic front end stays within a reasonable distance of reported single-robot numbers: De4D-SLAM reports 1.95 cm average ATE on Bonn, and Dream-SLAM reports 1.67 cm against ORB-SLAM3's 34.05 cm. This shows that moving-object tracks are reliable enough to be worth sending. On Swarm-SLAM's static multi-robot datasets (KITTI, GrAco, M2DGR, S3E), we confirm that preference-driven allocation does not hurt map accuracy when the operator cares only about geometry. In that case it should reduce to spectral prioritization.
Three findings would decide the matter. The first is a crossover: non-linear beats linear on threshold operators and ties on linear operators. Without that pattern, any win could be an artifact of tuning. The second is that the core/focus split preserves safety-zone reporting through a sudden refocus that exponential discounting forgets. The third is the bandwidth sweep showing where allocation stops mattering. If all three hold, the result is a concrete case that decentralized mapping under bandwidth limits is a preference-learning problem.
References
- Swarm-SLAM: Sparse Decentralized Collaborative Simultaneous Localization and Mapping Framework for Multi-Robot Systems
- Kimera-Multi: Robust, Distributed, Dense Metric-Semantic SLAM for Multi-Robot Systems
- De4D-SLAM: Gradient-Isolated Static-Dynamic Decoupling for Monocular SLAM in Dynamic Environments
- Dream-SLAM: Dreaming the Unseen for Active SLAM in Dynamic Environments
- Promptable Behaviors: Personalizing Multi-Objective Rewards from Human Preferences
- Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift
- RGB-D SLAM in Dynamic Environments Using Point Correlations
- BFTS: Thompson Sampling with Bayesian Additive Regression Trees
- Beyond Pairwise: Empowering LLM Alignment With (Ranked) Choice Modeling
- Beyond One-Preference-Fits-All Alignment: Multi-Objective Direct Preference Optimization
- Inferring Rewards from Language in Context
- LAMP: Large-Scale Autonomous Mapping and Positioning for Exploration of Perceptually-Degraded Subterranean Environments
- AcCoRD: Evaluating User-Agent Collaboration Under Realistic User Preference Dynamics
- GRADE: Generating Realistic and Dynamic Environments for robotics research with Isaac Sim