MoBaNet: Diverse Mobile Networking Systems

• MoBaNet is a polysemous term that defines mobility-aware systems, encompassing MPLS-based network management, bandit-driven access selection, multi-robot backbones, and multimodal segmentation frameworks.
• The architectures leverage techniques like hierarchical MPLS label binding, contextual bandits with blockchain integration, and optimization of wireless relay placement to enhance performance over dynamic networks.
• Recent developments include a parameter-efficient, modality-balanced fusion approach that integrates frozen Vision Foundation Models to reduce training costs while improving remote sensing semantic segmentation.

MoBaNet is not a single standardized term in the arXiv literature. It designates several unrelated technical systems: a Mobility Label Based Network (MLBN) for MPLS-based network-layer mobility management; a bandit-driven mobile access system for short-term mobile network provider selection; a Mobile Backbone Network for wireless infrastructure on demand in multi-robot teams; and, more recently, a parameter-efficient, modality-balanced fusion framework for multimodal remote sensing semantic segmentation. In adjacent usage, IMASNM is described as embodying the core principles typically associated with mobile-agent-based MoBaNet architectures, while “MoBaNet” is also explicitly distinguished from MoBiNet, the “Mobile Binary Network” for image classification (Berzin, 2021, Sandholm et al., 2020, Mox et al., 2020, Li et al., 18 Mar 2026, Sharma et al., 2012, Phan et al., 2019).

1. Terminological scope and principal usages

In the literature covered here, “MoBaNet” is best understood as a polysemous label rather than a single research program. The shared motif is mobility-aware or mobile-enabled networking, but the object of mobility differs sharply across papers: mobile users in MPLS cores, smartphone agents in spectrum markets, mobile relays in robotic ad-hoc networks, and multimodal feature streams in frozen Vision Foundation Model backbones. A separate but related branch uses mobile agents for scalable network management, and one source explicitly notes that the query term “MoBaNet” may be intended to refer to MoBiNet instead (Berzin, 2021, Sandholm et al., 2020, Mox et al., 2020, Li et al., 18 Mar 2026, Sharma et al., 2012, Phan et al., 2019).

Usage	Domain	Core definition
MoBaNet / MLBN	Network-layer mobility management	MPLS- and MP-BGP-based macro-/micro-mobility without Mobile IP
MoBaNet	Mobile access selection	Bandit-driven provider choice in a short-term spectrum market
MoBaNet	Multi-robot communications	Mobile backbone relays providing wireless infrastructure on demand
MoBaNet	Remote sensing segmentation	Parameter-efficient, modality-balanced symmetric fusion under frozen VFMs
IMASNM	Network management	Hierarchical mobile-agent architecture aligned with MoBaNet principles
MoBiNet	Image classification	Mobile Binary Network; distinct from MoBaNet

2. MoBaNet as Mobility Label Based Network (MLBN)

In "Mobility Label Based Network: Hierarchical Mobility Management and Packet Forwarding Architecture" (Berzin, 2021), MoBaNet, also referred to as MLBN, is a network-layer mobility management system that integrates the mobility control plane with the MPLS forwarding plane and relies solely on MPLS and Multi-Protocol BGP (MP-BGP). It eliminates Mobile IP elements and concepts such as Home Agents, Foreign Agents, Care-of-Addresses, and any L3 tunneling, yet provides both macro- and micro-mobility for IPv4/IPv6 mobile hosts and for network mobility via mobile routers. A Mobility Label is an overlay MPLS label bound to a mobile endpoint, while a Mobility Binding associates $$\{\text{mobile prefix}, \text{origin Router-ID}, \text{mobility label}\}$$. MLBN maps mobility onto pre-built, LDP-distributed, node-to-node infrastructure LSPs and distributes bindings through MP-BGP, thereby avoiding explicit per-mobile LSP setup, teardown, or redirection.

The architecture is hierarchical. Edge LERs with a Mobility Support Function terminate RAN connectivity, assign Mobility Labels, maintain per-mobile state, and originate Mobility Bindings. Area LERs allocate Local Mobility Labels and maintain FIB “label trails” that bind incoming and outgoing infrastructure labels to the Current Mobility Label. Area Mobility Route Reflectors store Mobility Bindings together with Internal and External Last Requestor Lists, implement on-demand distribution policies, and reflect mobility information within and across Mobility Areas. MLBN Border Edge Routers extend the mechanism to inter-carrier roaming. Micro-mobility includes MSF-Local handoff within a Mobility Region and Inter-MSF Intra-Area handoff within a Mobility Area; macro-mobility corresponds to Inter-MSF Inter-Area handoff and triggers scoped inter-area control updates and segmented LSP reselection.

The forwarding plane is strictly MPLS label-stack based. The typical stack depth is $d_{\text{stack}} \approx 2$: a top infrastructure LDP label toward a Next-Hop Router-ID and an inner mobility label, either a Local Mobility Label or a Current Mobility Label depending on the segment. At ALER boundaries, forwarding is implemented by MPLS pop, swap, and push operations using the label trail, with no IP prefix lookup. The control plane uses MP_REACH_NLRI and MP_UNREACH_NLRI with a mobility-specific NLRI format, selective downstream push, unsolicited downstream push, Area ID processing, and Last Requestor Lists to constrain state propagation to nodes that have actually consumed the binding.

The principal motivation is path optimality relative to Mobile IP. The paper models the penalty of triangular routing with a continuous-time Markov chain over hop-count differences and derives expected extra hops of $a = E[X(t)] = (k - 1)$ for fixed-to-mobile and $a = (3k - 2)/2$ for mobile-to-mobile traffic. With per-hop delay $\delta_i \sim N(\alpha,\sigma^2)$ and per-hop loss probability $p_i$, the cumulative penalties are modeled as $C_d \sim N(a\alpha, a\sigma^2)$ and $C_\ell = 1 - (1 - p_i)^a$. For $\alpha = 5$ ms, $\sigma = 2$ ms, $d_{\text{stack}} \approx 2$0, and $d_{\text{stack}} \approx 2$1, the additional one-way delay for mobile-to-mobile traffic is approximately $d_{\text{stack}} \approx 2$2–$d_{\text{stack}} \approx 2$3 ms and the success probability decreases by approximately $d_{\text{stack}} \approx 2$4. This establishes the analytical rationale for MLBN’s claim of optimal traffic delivery without Home-Agent anchoring.

3. MoBaNet as a bandit-driven mobile access market

In "A Multi-Armed Bandit-based Approach to Mobile Network Provider Selection" (Sandholm et al., 2020), MoBaNet is a bandit-driven mobile access system in which multiple operators expose short-term bandwidth leases in an open spectrum market, and an autonomous on-device agent selects the provider that maximizes user utility subject to price and budget. Providers list frequency band, bandwidth, epoch, price per allocation, and maximum allocations per epoch; the user’s smartphone agent observes context $d_{\text{stack}} \approx 2$5, where the simplest deployed contextual formulation reduces to $d_{\text{stack}} \approx 2$6. The step reward is defined as value per price, $d_{\text{stack}} \approx 2$7, with a budget-aware form that multiplies the utility by remaining-budget factors. The objective is to maximize expected cumulative utility, and the paper also states regret relative to the optimal policy as $d_{\text{stack}} \approx 2$8.

The deployed policies are lightweight. The contextual Monte Carlo $d_{\text{stack}} \approx 2$9-armed bandit uses $a = E[X(t)] = (k - 1)$0-greedy action selection with a running reward estimate $a = E[X(t)] = (k - 1)$1, short moving windows, and optional UCB or softmax variants. Tabular Q-learning is also implemented with update

$a = E[X(t)] = (k - 1)$2

but the reported results favor the contextual bandit with a two-period history window. The paper discusses Gittins indices and restless-bandit extensions for asynchronous prices, yet direct contextual bandits and tabular Q-learning are the deployed mechanisms.

The system architecture couples learning with market execution. The ledger is implemented with Hyperledger Sawtooth and a custom transaction processor supporting offer, allocate, deposit, and withdraw. On-device execution uses GSMA SGP.22-compliant eSIM provisioning and a privileged Profile Manager Service that switches pre-installed profiles. The Android agent comprises a Management UI, Monitor Broadcast Receiver, Location Service, LTE and app-usage monitors, a Statistics Normalizer, a Provider Predictor, and the Profile Manager Service. In the LTE prototype, custom control messages and, in one integration, PBCH/SIB extensions expose offer metadata; the measured overhead is approximately $a = E[X(t)] = (k - 1)$3 s for the NAS path and approximately $a = E[X(t)] = (k - 1)$4 s to reinitialize the srsLTE UE with a new IMSI.

Empirical results are reported for both a PyLTEs-based simulator and a real LTE testbed. In simulation, the typical QoE or utility improvement relative to Random is in the $a = E[X(t)] = (k - 1)$5–$a = E[X(t)] = (k - 1)$6 range depending on price, demand, mobility, and competition. Specific gains include approximately $a = E[X(t)] = (k - 1)$7 over Random under variable location with fixed price and approximately $a = E[X(t)] = (k - 1)$8 under variable location with variable price. In the LTE testbed, the agent records $a = E[X(t)] = (k - 1)$9–$a = (3k - 2)/2$0 QoE gains across competing-agent scenarios, and training is sample-efficient: after $a = (3k - 2)/2$1 training samples, allocation success reaches approximately $a = (3k - 2)/2$2 and utility reaches approximately $a = (3k - 2)/2$3 of optimal. The paper’s central claim is therefore not merely algorithmic; it is an end-to-end demonstration combining contextual bandits, blockchain transaction flow, and eSIM-based provider switching.

4. MoBaNet as a mobile backbone for multi-robot wireless infrastructure

In "Mobile Wireless Network Infrastructure on Demand" (Mox et al., 2020), MoBaNet is instantiated as a Mobile Backbone Network in which a dedicated set of mobile robots, the “network team,” provides and continuously reconfigures an ad-hoc wireless backbone to satisfy the end-to-end communication requirements of a separate “task team.” The task team acts as the client edge and specifies per-flow rate and confidence requirements; the network team acts as the mobile backbone and plans both routing and relay placement. This explicit separation decouples mission planning from communication planning and produces a task-agnostic service architecture.

The communication model is a directed, weighted multi-hop graph over all task and network agents, with stochastic link rate $a = (3k - 2)/2$4, single-radio half-duplex operation, and per-node transmit and receive time-sharing constraints. The paper gives both a general physical-layer interpretation using path loss and SINR and a normalized probabilistic rate model. The mean rate is

$a = (3k - 2)/2$5

and the variance term is

$a = (3k - 2)/2$6

with parameters $a = (3k - 2)/2$7 dBm, $a = (3k - 2)/2$8 dBm, $a = (3k - 2)/2$9, $\delta_i \sim N(\alpha,\sigma^2)$0, and $\delta_i \sim N(\alpha,\sigma^2)$1. End-to-end requirements are encoded as chance constraints on node-wise flow balances.

Given positions $\delta_i \sim N(\alpha,\sigma^2)$2, routing is posed as a convex SOCP over time-sharing variables $\delta_i \sim N(\alpha,\sigma^2)$3 and a robustness slack $\delta_i \sim N(\alpha,\sigma^2)$4. The flow balance mean and variance are

$\delta_i \sim N(\alpha,\sigma^2)$5

$\delta_i \sim N(\alpha,\sigma^2)$6

and the chance constraint is written as

$\delta_i \sim N(\alpha,\sigma^2)$7

Relay motion is then handled by a sample-based local controller that seeks to improve the worst-case feasibility margin

$\delta_i \sim N(\alpha,\sigma^2)$8

where larger $\delta_i \sim N(\alpha,\sigma^2)$9 indicates greater headroom. The algorithm alternates between SOCP routing updates and local search over collision-free relay configurations.

The simulation study uses three task agents patrolling a circle of radius $p_i$0 m with three multicast flows and required source-node margin $p_i$1 at confidence $p_i$2. For $p_i$3 relay, both fixed and dynamic configurations converge to the center and achieve approximately $p_i$4 normalized margin, which does not meet the specification. For $p_i$5, a fixed triangle fails intermittently, while the dynamic triangle rotates with the task team and consistently meets the rate and confidence specifications. For $p_i$6, both fixed and dynamic teams satisfy the specifications, with dynamic relay placement outperforming on average. Hardware validation on Intel Aero quadrotors, Linux/ROS, IEEE 802.11n IBSS, and a centralized planner running at approximately $p_i$7 Hz shows that, as the roaming task agent moves away from the base, the direct-link baseline suffers throughput drop and latency increase, whereas probabilistic routing via the mobile relay keeps throughput much more stable and delay bounded even after the direct link breaks.

5. MoBaNet as a parameter-efficient multimodal fusion framework

In "Parameter-Efficient Modality-Balanced Symmetric Fusion for Multimodal Remote Sensing Semantic Segmentation" (Li et al., 18 Mar 2026), MoBaNet is a multimodal segmentation framework that fuses complementary remote sensing modalities with a largely frozen Vision Foundation Model backbone. It addresses two stated bottlenecks: the training and storage cost of full fine-tuning, and modality imbalance, defined as the tendency during optimization for the model to rely predominantly on the modality that is easiest to learn, often RGB, thereby suppressing auxiliary modalities such as DSM. The framework uses a symmetric dual-stream design with frozen Transformer blocks and introduces trainable modules only where cross-modal interaction and adaptive fusion are required.

The architecture is organized around three components. The Cross-modal Prompt-Injected Adapter (CPIA) performs pre-stage token-level interaction. Given paired stage inputs $p_i$8 and $p_i$9, the Cross-modal Prompt Generator projects both to reduced dimension, fuses them into a shared semantic base $C_d \sim N(a\alpha, a\sigma^2)$0, and applies Task Feature Transformation:

$C_d \sim N(a\alpha, a\sigma^2)$1

The prompt-conditioned adapter is

$C_d \sim N(a\alpha, a\sigma^2)$2

The Difference-Guided Gated Fusion Module (DGFM) operates after each selected stage. It reduces both modality features, computes discrepancy $C_d \sim N(a\alpha, a\sigma^2)$3, predicts a gate

$C_d \sim N(a\alpha, a\sigma^2)$4

and fuses full-channel features by

$C_d \sim N(a\alpha, a\sigma^2)$5

Finally, Modality-Conditional Random Masking (MCRM) is a training-only strategy that masks one modality per selected sample and applies hard-pixel auxiliary supervision on the error set $C_d \sim N(a\alpha, a\sigma^2)$6, with total loss $C_d \sim N(a\alpha, a\sigma^2)$7.

The framework supports DINOv2 ViT-B and ViT-L as well as SAM ViT-B and ViT-L. The frozen components are the backbone self-attention, MLP, optical patch embedding, and shared absolute positional encoding; the trainable components are the auxiliary patch embedding, CPIA, DGFM, UperNet decoder, and lightweight auxiliary heads used only during training. With a DINOv2 ViT-B backbone, the trainable parameter count is $C_d \sim N(a\alpha, a\sigma^2)$8M, compared with $C_d \sim N(a\alpha, a\sigma^2)$9M for ABCNet, $C_\ell = 1 - (1 - p_i)^a$0M for UNetFormer, $C_\ell = 1 - (1 - p_i)^a$1M for DC-Swin, $C_\ell = 1 - (1 - p_i)^a$2M for FTransUNet, and $C_\ell = 1 - (1 - p_i)^a$3M for MANet.

Evaluation is performed on ISPRS Vaihingen and Potsdam. The best reported Vaihingen result is obtained by MoBaNet with DINOv2 ViT-L, reaching OA $C_\ell = 1 - (1 - p_i)^a$4, mF1 $C_\ell = 1 - (1 - p_i)^a$5, and mIoU $C_\ell = 1 - (1 - p_i)^a$6. The best Potsdam result is also achieved by MoBaNet with DINOv2 ViT-L, reaching OA $C_\ell = 1 - (1 - p_i)^a$7, mF1 $C_\ell = 1 - (1 - p_i)^a$8, and mIoU $C_\ell = 1 - (1 - p_i)^a$9; the DINOv2 ViT-B variant reaches OA $\alpha = 5$0, mF1 $\alpha = 5$1, and mIoU $\alpha = 5$2. The DINOv2 ViT-B configuration reports latency $\alpha = 5$3 ms, FPS $\alpha = 5$4, and memory $\alpha = 5$5 MB. Robustness to missing modality is a central result: on Potsdam with RGB-only inference and DSM set to zero, the performance drop is OA $\alpha = 5$6, mF1 $\alpha = 5$7, and mIoU $\alpha = 5$8, whereas a reference baseline without MCRM suffers a much larger mIoU drop of $\alpha = 5$9. Ablations on Potsdam show a progression from a base model with $\sigma = 2$0M trainable parameters and mIoU $\sigma = 2$1, to $\sigma = 2$2 at $\sigma = 2$3, $\sigma = 2$4 at $\sigma = 2$5, and the full model at $\sigma = 2$6.

6. Related names, adjacent architectures, and common confusions

A related but distinct line appears in "An Intelligent Mobile-Agent Based Scalable Network Management Architecture for Large-Scale Enterprise System" (Sharma et al., 2012). The paper does not use the term “MoBaNet” explicitly, but it is described as embodying the core principles typically associated with mobile-agent-based MoBaNet architectures: mobility, local autonomy, hierarchical delegation, and lightweight inter-manager communication. Its entities are the Global Network Manager, Mobile Subnetwork Layer Managers, Mobile Data Agents, and provisioning or event-reporting deglets. The cost model contrasts centralized SNMP/CMIP polling with hierarchical mobile-agent management. For one poll over five variables per node, the reported centralized cost is approximately $\sigma = 2$7 KB, whereas IMASNM reports approximately $\sigma = 2$8 KB; for $\sigma = 2$9 polls per hour, the reported costs are $d_{\text{stack}} \approx 2$00 bytes versus $d_{\text{stack}} \approx 2$01 bytes, a reduction of approximately $d_{\text{stack}} \approx 2$02. In this sense, IMASNM is best treated as an adjacent MoBaNet-style network-management architecture rather than as one of the principal MoBaNet definitions.

A separate nomenclature issue concerns MoBiNet. "MoBiNet: A Mobile Binary Network for Image Classification" (Phan et al., 2019) is not a MoBaNet paper, although one source explicitly states that the query’s “MoBaNet” refers to MoBiNet. MoBiNet binarizes a MobileNet-derived backbone, introduces skip connections and extra binary $d_{\text{stack}} \approx 2$03 pointwise layers to stabilize binary depthwise-separable training, replaces ReLU with PReLU, and trains from scratch without float pretraining. On ImageNet it achieves up to $d_{\text{stack}} \approx 2$04 Top-1 and $d_{\text{stack}} \approx 2$05 Top-5, with reported model size approximately $d_{\text{stack}} \approx 2$06–$d_{\text{stack}} \approx 2$07 MB and markedly lower FLOPs than full-precision MobileNet and several binary ResNet baselines. The overlap is therefore lexical rather than conceptual.

Taken together, the literature shows that “MoBaNet” functions as a cross-domain label rather than a canonical architecture. In networking, it most often refers either to MPLS-based hierarchical mobility management or to user-centric, learning-based mobile access selection. In robotics, it denotes a mobile backbone that optimizes end-to-end flow guarantees under stochastic links. In remote sensing, it denotes a modality-balanced PEFT framework over frozen VFMs. Any technical use of the term therefore requires immediate disambiguation by domain, control model, and cited source.