MAVN: Adaptive Nodes & Navigation

• MAVN is an umbrella term that encompasses adaptive virtual nodes for dynamic message passing on graphs, multi-agent visual navigation, and micro aerial vehicle navigation in challenging environments.
• It employs adaptive connectivity and learned communication strategies to optimize system performance in diverse applications ranging from embodied AI to aerial robotics.
• Its practical implications span improved graph neural network flexibility, enhanced cooperative navigation in 3D settings, and robust operation in GNSS-denied or degraded sensing scenarios.

Searching arXiv for the acronym and closely related titles to ground the article in the relevant papers. arXiv search query: MAVN MAVEN quadrotor graph virtual nodes visual navigation

MAVN is an overloaded acronym in recent arXiv literature. In the provided corpus, it denotes an adaptive virtual-node framework for message passing on graphs, Multi-Agent Visual Navigation in photorealistic 3D environments, and Micro Aerial Vehicle Navigation in GNSS-denied or perceptually degraded physical settings. Closely related names—notably MAVEN and MAVNet—refer to distinct systems in affective computing, agentic video annotation, agile quadrotor meta-reinforcement learning, and road-following UAV imitation learning. Disambiguation is therefore a prerequisite for correct citation, reproduction, and comparison (Lee et al., 2 Jun 2026, Wang et al., 2021, Kominiak et al., 2020, Ahire et al., 16 Mar 2025, Zhang et al., 21 May 2026, Zhou et al., 11 Mar 2026, Kumaar et al., 2018).

1. Scope and terminological disambiguation

In the provided literature, the string “MAVN” is used in multiple non-equivalent senses. Some usages denote a named method, while others denote a task family or a broader problem domain.

Usage in the provided literature	Meaning	Representative paper
MAVN	Adaptive virtual nodes for dynamic message passing on graphs	(Lee et al., 2 Jun 2026)
MAVN	Multi-Agent Visual Navigation	(Wang et al., 2021)
MAVN	Micro Aerial Vehicle Navigation	(Kominiak et al., 2020, Schleich et al., 2021, Mansouri et al., 2020, Yadav et al., 2022, Bähnemann et al., 2021, Kulkarni et al., 2024)
Related but distinct	MAVEN / MAVNet	(Ahire et al., 16 Mar 2025, Zhang et al., 21 May 2026, Zhou et al., 11 Mar 2026, Kumaar et al., 2018)

This terminological spread has concrete consequences. In graph learning, MAVN is an end-to-end differentiable MPNN framework with adaptive virtual nodes (Lee et al., 2 Jun 2026). In embodied AI, MAVN denotes a cooperative visual navigation setting formalized as a Dec-POMDP and benchmarked through CollaVN (Wang et al., 2021). In robotics, MAVN is used as shorthand for navigation by micro aerial vehicles in underground tunnels, disaster sites, or cluttered local-sensing regimes (Kominiak et al., 2020, Schleich et al., 2021, Mansouri et al., 2020, Yadav et al., 2022, Bähnemann et al., 2021, Kulkarni et al., 2024).

2. MAVN as adaptive virtual nodes in graph neural networks

The most literal paper-title use of MAVN is "Learn When and Where to Connect: Adaptive Virtual Nodes for Dynamic Message Passing on Graphs" (Lee et al., 2 Jun 2026). Here MAVN is an architecture-agnostic augmentation to MPNNs that learns both when to instantiate virtual nodes and where to connect them, layer by layer, from a pool of candidate VNs. The method is motivated by three limitations of prior VN-based schemes: they often constrain all nodes to connect to the same number of VNs, fix the connections before message passing, and decide node-to-VN edges independently of the other nodes sharing a VN.

At layer $l$, MAVN computes node–VN relevance scores

$$s_{vz}^{(l)} = \mathrm{MLP}^{(l)}(x_{v}^{(l-1)}) \cdot k_{z}^{(l)} / \sqrt{d_{\mathrm{dot}}}$$

and then adjusts them by a global competition term

$\bar{s}_{vz}^{(l)} = s_{vz}^{(l)} + \alpha \cdot \logsoftmax(S^{(l)})[s_{vz}^{(l)}].$

Candidate VNs are selected through a smooth-max aggregation

$s_{z}^{(l)} = \logmeanexp(\{\bar{s}_{vz}^{(l)} \mid v \in V^{(l-1)}\}),$

with a threshold rule $\sigma(s_z^{(l)}) \ge 0.5$. Node–VN connectivity is then decided through a dual-perspective score that combines the node’s preference for a VN and the VN’s preference over nodes, weighted by a learned $\beta^{(l)}$. Selected VNs receive representations through a seed–aggregation mixture

$$x_{z}^{(l-1)} = \gamma^{(l)} \odot q_{z}^{(l)} + (1-\gamma^{(l)}) \odot \left(\sum_{v \in N_z^{(l)}} p_{vz}^{(l)} x_v^{(l-1)} / c_z^{(l)}\right),$$

after which the backbone MPNN runs on the augmented graph.

The paper proves a strong flexibility result: for any node–VN connectivity pattern, there exists a parameter configuration of a single-layer MAVN that can simulate that pattern. Under the theorem’s distinguishability assumption on node representations, MAVN can realize arbitrary subsets $V_1,\dots,V_K$ by introducing $K$ virtual nodes and connecting each $z_i$ exactly to $$s_{vz}^{(l)} = \mathrm{MLP}^{(l)}(x_{v}^{(l-1)}) \cdot k_{z}^{(l)} / \sqrt{d_{\mathrm{dot}}}$$0 (Lee et al., 2 Jun 2026). This gives a formal sense in which MAVN subsumes fixed and fully connected VN schemes.

Empirically, experiments on nine real-world datasets show consistent gains over backbone MPNNs. Reported improvements reach up to $$s_{vz}^{(l)} = \mathrm{MLP}^{(l)}(x_{v}^{(l-1)}) \cdot k_{z}^{(l)} / \sqrt{d_{\mathrm{dot}}}$$1 over the backbones, including a PascalVOC-SP increase from $$s_{vz}^{(l)} = \mathrm{MLP}^{(l)}(x_{v}^{(l-1)}) \cdot k_{z}^{(l)} / \sqrt{d_{\mathrm{dot}}}$$2 to $$s_{vz}^{(l)} = \mathrm{MLP}^{(l)}(x_{v}^{(l-1)}) \cdot k_{z}^{(l)} / \sqrt{d_{\mathrm{dot}}}$$3 F1-macro for MAVN-GCN, and statistical significance at $$s_{vz}^{(l)} = \mathrm{MLP}^{(l)}(x_{v}^{(l-1)}) \cdot k_{z}^{(l)} / \sqrt{d_{\mathrm{dot}}}$$4 in $$s_{vz}^{(l)} = \mathrm{MLP}^{(l)}(x_{v}^{(l-1)}) \cdot k_{z}^{(l)} / \sqrt{d_{\mathrm{dot}}}$$5 one-tailed paired $$s_{vz}^{(l)} = \mathrm{MLP}^{(l)}(x_{v}^{(l-1)}) \cdot k_{z}^{(l)} / \sqrt{d_{\mathrm{dot}}}$$6-tests on heterophilic graphs (Lee et al., 2 Jun 2026). Ablations further show that layer-wise adaptivity, selective sparsity, dual-perspective scoring, and data-dependent VN states are all important; introducing VNs only at the first layer, connecting all nodes to all selected VNs, or collapsing the two perspectives degrades performance.

3. MAVN as Multi-Agent Visual Navigation

In "Collaborative Visual Navigation," MAVN denotes Multi-Agent Visual Navigation, a cooperative embodied-AI setting in which multiple agents navigate photorealistic 3D environments to reach target locations (Wang et al., 2021). The problem is formalized as an $$s_{vz}^{(l)} = \mathrm{MLP}^{(l)}(x_{v}^{(l-1)}) \cdot k_{z}^{(l)} / \sqrt{d_{\mathrm{dot}}}$$7-agent extension of partially observable Markov decision processes, with private observations $$s_{vz}^{(l)} = \mathrm{MLP}^{(l)}(x_{v}^{(l-1)}) \cdot k_{z}^{(l)} / \sqrt{d_{\mathrm{dot}}}$$8, stochastic policies $$s_{vz}^{(l)} = \mathrm{MLP}^{(l)}(x_{v}^{(l-1)}) \cdot k_{z}^{(l)} / \sqrt{d_{\mathrm{dot}}}$$9, transition function $\bar{s}_{vz}^{(l)} = s_{vz}^{(l)} + \alpha \cdot \logsoftmax(S^{(l)})[s_{vz}^{(l)}].$0, and per-agent returns $\bar{s}_{vz}^{(l)} = s_{vz}^{(l)} + \alpha \cdot \logsoftmax(S^{(l)})[s_{vz}^{(l)}].$1. The environment provides first-person panoramic RGB observations of shape $\bar{s}_{vz}^{(l)} = s_{vz}^{(l)} + \alpha \cdot \logsoftmax(S^{(l)})[s_{vz}^{(l)}].$2, goal images $\bar{s}_{vz}^{(l)} = s_{vz}^{(l)} + \alpha \cdot \logsoftmax(S^{(l)})[s_{vz}^{(l)}].$3, and continuous wheel-velocity actions $\bar{s}_{vz}^{(l)} = s_{vz}^{(l)} + \alpha \cdot \logsoftmax(S^{(l)})[s_{vz}^{(l)}].$4.

The benchmark is CollaVN, built on iGibsonV1 with 572 full buildings covering 211,000 m$\bar{s}_{vz}^{(l)} = s_{vz}^{(l)} + \alpha \cdot \logsoftmax(S^{(l)})[s_{vz}^{(l)}].$5. It uses Gibson-tiny splits of 25 train, 5 val, and 5 test scenes, and contains 1M train, 60K val, and 120K test episodes across its sub-datasets (Wang et al., 2021). Three task variants are defined. In CommonGoal, all agents share the same goal image. In SpecificGoal, agents pursue different goals. In Ad-HoCoop, team size changes between training and testing, such as $\bar{s}_{vz}^{(l)} = s_{vz}^{(l)} + \alpha \cdot \logsoftmax(S^{(l)})[s_{vz}^{(l)}].$6 or $\bar{s}_{vz}^{(l)} = s_{vz}^{(l)} + \alpha \cdot \logsoftmax(S^{(l)})[s_{vz}^{(l)}].$7.

The central methodological contribution is a memory-augmented communication framework. Each agent maintains a learned local map $\bar{s}_{vz}^{(l)} = s_{vz}^{(l)} + \alpha \cdot \logsoftmax(S^{(l)})[s_{vz}^{(l)}].$8 and a private external memory $\bar{s}_{vz}^{(l)} = s_{vz}^{(l)} + \alpha \cdot \logsoftmax(S^{(l)})[s_{vz}^{(l)}].$9 that stores past communication features. Communication uses learned query, key, and value modules with $s_{z}^{(l)} = \logmeanexp(\{\bar{s}_{vz}^{(l)} \mid v \in V^{(l-1)}\}),$0 and $s_{z}^{(l)} = \logmeanexp(\{\bar{s}_{vz}^{(l)} \mid v \in V^{(l-1)}\}),$1. A requester broadcasts $s_{z}^{(l)} = \logmeanexp(\{\bar{s}_{vz}^{(l)} \mid v \in V^{(l-1)}\}),$2; supporting agents score it against their current and past keys, aggregate weighted values, and pass through an activation $s_{z}^{(l)} = \logmeanexp(\{\bar{s}_{vz}^{(l)} \mid v \in V^{(l-1)}\}),$3 with $s_{z}^{(l)} = \logmeanexp(\{\bar{s}_{vz}^{(l)} \mid v \in V^{(l-1)}\}),$4. If $s_{z}^{(l)} = \logmeanexp(\{\bar{s}_{vz}^{(l)} \mid v \in V^{(l-1)}\}),$5, no communication occurs at that step. This implements both who to communicate with and when to communicate.

Training uses fully decentralized PPO actor-critic with $s_{z}^{(l)} = \logmeanexp(\{\bar{s}_{vz}^{(l)} \mid v \in V^{(l-1)}\}),$6, entropy coefficient $s_{z}^{(l)} = \logmeanexp(\{\bar{s}_{vz}^{(l)} \mid v \in V^{(l-1)}\}),$7, value loss coefficient $s_{z}^{(l)} = \logmeanexp(\{\bar{s}_{vz}^{(l)} \mid v \in V^{(l-1)}\}),$8, Adam with learning rate $s_{z}^{(l)} = \logmeanexp(\{\bar{s}_{vz}^{(l)} \mid v \in V^{(l-1)}\}),$9, $\sigma(s_z^{(l)}) \ge 0.5$0 mini-batches, and $\sigma(s_z^{(l)}) \ge 0.5$1 epochs per update on $\sigma(s_z^{(l)}) \ge 0.5$2 NVIDIA Tesla V100 32GB (Wang et al., 2021). The paper introduces Success weighted by Step Ratio,

$\sigma(s_z^{(l)}) \ge 0.5$3

to complement SR, DTS, and SPL.

Across CommonGoal, SpecificGoal, and Ad-Hoc settings, the memory-augmented method consistently outperforms imitation learning, MARL without communication, and MARL without memory. For CommonGoal with $\sigma(s_z^{(l)}) \ge 0.5$4, the reported overall results are SR 23.10, DTS 2.90, SSR 1.45, SPL 0.11, compared with SR 18.70, DTS 3.12, SSR 1.21, SPL 0.09 for the memoryless variant and SR 17.72, DTS 3.17, SSR 1.12, SPL 0.09 for the no-communication variant (Wang et al., 2021). The same ranking persists under heterogeneous goals and ad-hoc team-size changes.

4. MAVN as Micro Aerial Vehicle Navigation

A third usage treats MAVN as Micro Aerial Vehicle Navigation, especially in GNSS-denied, dark, dusty, or cluttered environments. The provided literature spans low-cost mining platforms, subterranean NMPC, LiDAR-centric disaster-response autonomy, reactive target interception, and high-precision localization for airborne ground-penetrating radar (Kominiak et al., 2020, Schleich et al., 2021, Mansouri et al., 2020, Yadav et al., 2022, Bähnemann et al., 2021, Kulkarni et al., 2024).

A representative underground-mining system is a low-cost, modular, and consumable quadrotor based on an Enzo330 V2 frame, ROSFlight on an AfroFlight NAZE32 Rev6, and an Aaeon UP-Board companion computer (Kominiak et al., 2020). Its sensor suite includes an RPLidar A2M8, LIDAR-Lite 3, PX4FLOW, and a PlayStation 3 Eye camera. The platform uses potential-field obstacle avoidance, tunnel-axis following, online 2D occupancy mapping at ~1 Hz and 0.05 m/pixel, and PANOC-based NMPC. It reports flight time ≈ 12 min, illumination at 1 m of 2200 lux at maximum power, 460 lux at 1 m in the mine trial, and total cost ≈ \$1,095 USD (Kominiak et al., 2020).

For autonomous flight in unknown GNSS-denied disaster environments, another system uses a DJI Matrice 210 v2, an Intel NUC8i7BEH, and an Ouster OS-0 3D LiDAR (Schleich et al., 2021). LiDAR odometry runs at 10 Hz, EKF state estimation at 50 Hz, occupancy mapping at 10 Hz, planning at 1 Hz, and MPC at 50 Hz. The occupancy map uses 25 cm voxels and removes measurements older than $\sigma(s_z^{(l)}) \ge 0.5$5 scans. Demonstrations include outdoor facade inspection, indoor–outdoor transitions, and precise hovering without GNSS, with maximum flight speed capped at 1 m/s for safety near structures.

Subterranean NMPC is further specialized in "Subterranean MAV Navigation based on Nonlinear MPC with Collision Avoidance Constraints" (Mansouri et al., 2020). The MAV is modeled as a floating object that tracks $\sigma(s_z^{(l)}) \ge 0.5$6, $\sigma(s_z^{(l)}) \ge 0.5$7, and altitude $\sigma(s_z^{(l)}) \ge 0.5$8, while collision constraints are derived from 2D LiDAR distances $\sigma(s_z^{(l)}) \ge 0.5$9. The field-tested configuration uses horizon $\beta^{(l)}$0, sampling time $\beta^{(l)}$1, minimum clearance $\beta^{(l)}$2, and reports average solve time ≈ 10 ms on an Aaeon UP-Board. In the reported trials, reference speeds reached 1.2 m/s and no collisions occurred.

A more aggressive local-sensing formulation appears in "Receding Horizon Navigation and Target Tracking for Aerial Detection of Transient Radioactivity" (Yadav et al., 2022). This system uses an Intel RealSense D435 RGB-D camera at 30 Hz, a T265 visual–inertial sensor with OpenVINS at ~30 Hz, and an onboard SSD-MobileNetV2 detector at ~15 Hz. Candidate trajectories are generated within a camera-centered pyramid FOV and ranked by endpoint-to-target distance and a collision cost $\beta^{(l)}$3. The planner produces safe, dynamically feasible trajectories without a global planner or prior map, and experimental speeds reach 4.5–5 m/s in cluttered indoor and outdoor environments.

At the precision end of MAVN, "Under the Sand" presents an airborne GPSAR system on a DJI M600 Pro with dual RTK GNSS, industrial-grade IMU, radar altimeter, lidar altimeter, and FMCW GPR (Bähnemann et al., 2021). The system reports sensor timing accuracy ≈ 0.8 $\beta^{(l)}$4s, timing precision ≈ 0.05 $\beta^{(l)}$5s with 0.1 $\beta^{(l)}$6s clock resolution, and localization rates of 1 kHz. The dual-position factor formulation improves online localization accuracy by up to 40% and batch localization accuracy by up to 59% relative to a single position factor with uncertain heading initialization. These margins are sufficient for coherent radar backprojection and field detection of shallow buried targets.

The broader review "Aerial Field Robotics" places such systems in a resilience-oriented autonomy stack (Kulkarni et al., 2024). Across the chapter, MAV navigation is framed by collision tolerance, sensing degradation, constrained compute, health-aware fusion, ESDF/TSDF mapping, receding-horizon planning, and mode-based fallbacks. This suggests that, in robotics usage, MAVN is less a single algorithm than a family of design patterns for operating small aerial vehicles where contact, drift, darkness, dust, or degraded perception are expected rather than exceptional.

5. Related but distinct terms: MAVEN and MAVNet

Several nearby names are easily confused with MAVN but refer to different systems. In affective computing, "MAVEN: Multi-modal Attention for Valence-Arousal Emotion Network" integrates visual, audio, and textual modalities through six directed cross-modal attention pathways and predicts polar affect $\beta^{(l)}$7, later converted by $\beta^{(l)}$8 and $\beta^{(l)}$9 (Ahire et al., 16 Mar 2025). On Aff-Wild2, it reports CCC$$x_{z}^{(l-1)} = \gamma^{(l)} \odot q_{z}^{(l)} + (1-\gamma^{(l)}) \odot \left(\sum_{v \in N_z^{(l)}} p_{vz}^{(l)} x_v^{(l-1)} / c_z^{(l)}\right),$$0, CCC$$x_{z}^{(l-1)} = \gamma^{(l)} \odot q_{z}^{(l)} + (1-\gamma^{(l)}) \odot \left(\sum_{v \in N_z^{(l)}} p_{vz}^{(l)} x_v^{(l-1)} / c_z^{(l)}\right),$$1, and average CCC = 0.3543, compared with a cited baseline average of 0.201.

A second MAVEN is "A Multi-stage Agentic Video Event aNnotation" (Zhang et al., 21 May 2026). That pipeline uses three-stage captioning and consolidation into an explicit Multi-Scale Spatio-Temporal Event Description $$x_{z}^{(l-1)} = \gamma^{(l)} \odot q_{z}^{(l)} + (1-\gamma^{(l)}) \odot \left(\sum_{v \in N_z^{(l)}} p_{vz}^{(l)} x_v^{(l-1)} / c_z^{(l)}\right),$$2, then generates MCQ, binary verification, and open-ended QA from MSTED alone. It labels over 5,300 traffic videos and fine-tunes Cosmos-Reason2-8B. On a private CCTV set, CR2 + CCTV SFT reaches 86.25 MCQ, 85.00 Verif, 39.45 Open, and + RL reaches 88.75, 81.25, 37.29; on AccidentBench, the final + RL system reaches 44.2 overall MCQ, surpassing the reported Gemini baselines (Zhang et al., 21 May 2026).

A third MAVEN is a meta-reinforcement-learning framework for agile quadrotor maneuvers under varying dynamics (Zhou et al., 11 Mar 2026). It uses a predictive context encoder $$x_{z}^{(l-1)} = \gamma^{(l)} \odot q_{z}^{(l)} + (1-\gamma^{(l)}) \odot \left(\sum_{v \in N_z^{(l)}} p_{vz}^{(l)} x_v^{(l-1)} / c_z^{(l)}\right),$$3, latent dimension $$x_{z}^{(l-1)} = \gamma^{(l)} \odot q_{z}^{(l)} + (1-\gamma^{(l)}) \odot \left(\sum_{v \in N_z^{(l)}} p_{vz}^{(l)} x_v^{(l-1)} / c_z^{(l)}\right),$$4, and PPO-conditioned control policy $$x_{z}^{(l-1)} = \gamma^{(l)} \odot q_{z}^{(l)} + (1-\gamma^{(l)}) \odot \left(\sum_{v \in N_z^{(l)}} p_{vz}^{(l)} x_v^{(l-1)} / c_z^{(l)}\right),$$5. In real-world tests it adapts online to mass increases of up to 66.7% and single-rotor thrust losses as severe as 70%, with onboard inference at 100 Hz and training convergence in less than an hour through GPU-vectorized simulation (Zhou et al., 11 Mar 2026).

MAVNet, finally, is a road-following UAV imitation-learning model rather than MAVN proper (Kumaar et al., 2018). It is a 39-layer Inception-style network operating on 100 × 100 tomographic reconstructions, runs at about 30 FPS on a CPU-only Intel Core i3 laptop, reports 98.44% accuracy, and demonstrates a 357 m continuous autonomous stretch without crashing or overshooting (Kumaar et al., 2018).

6. Cross-cutting patterns and interpretive synthesis

Across the provided literature, MAVN and its near neighbors occupy very different technical domains, but several recurrent design choices appear. One is adaptive connectivity: MAVN on graphs decides layer-wise VN instantiation and edge formation; Multi-Agent Visual Navigation decides who communicates and when; the meta-RL MAVEN infers latent dynamics online from context; and several micro-aerial systems adapt planning or control from local sensor evidence rather than fixed global structure (Lee et al., 2 Jun 2026, Wang et al., 2021, Zhou et al., 11 Mar 2026, Mansouri et al., 2020).

A second recurring theme is the use of an explicit intermediate state that stabilizes downstream reasoning or control. Examples include MSTED in video annotation, local maps and private memories in CollaVN, factor-graph states in GPSAR localization, and occupancy or ESDF-style spatial representations in aerial robotics (Zhang et al., 21 May 2026, Wang et al., 2021, Bähnemann et al., 2021, Kulkarni et al., 2024). This suggests that many systems bearing the MAVN/MAVEN family of names are not purely end-to-end, even when learned components dominate.

A third pattern is that resource constraints remain method-defining rather than incidental. The low-cost mine platform is designed around replaceability and operator data access; disaster-response navigation is organized around onboard rates and safe replanning; subterranean NMPC is shaped by embedded solve times; CollaVN agents communicate under low-bandwidth assumptions; and adaptive virtual nodes are motivated partly by improving long-range message passing without fully dense rewiring (Kominiak et al., 2020, Schleich et al., 2021, Mansouri et al., 2020, Wang et al., 2021, Lee et al., 2 Jun 2026).

A plausible implication is that acronym-only retrieval on arXiv is unreliable for this family of terms. In the provided literature, “MAVN” can point to graph rewiring, cooperative visual navigation, or aerial robot autonomy, while “MAVEN” can denote emotion recognition, agentic annotation, or meta-RL for quadrotors. For technical work, the operative identifier is therefore not the acronym alone but the full expansion, domain, and arXiv id.