AI Routing Framework Learns From Five Sensor Streams, But Classical Planners Still Win

Researchers have built an end-to-end artificial intelligence framework, called MARLIN, that learns delivery-delay prediction and adaptive routing directly from five raw Internet-of-Things sensor streams: temporal traffic sequences, LiDAR point-cloud features, camera embeddings, vehicle telemetry, and weather data. The system, described in the journal Discover Internet of Things, combines reconstruction-based pre-training of modality-specific encoders, gated cross-attention fusion, an SLA-aware prediction head, and a graph-based Soft Actor-Critic reinforcement-learning agent. Yet the study is equally notable for how carefully it limits its own claims: the results come from a controlled benchmark in which the delivery-delay labels were generated by a physics-informed mathematical model rather than observed from real logistics operations.

The benchmark draws on real data from four public sources. NYC Taxi trip records from January 2023 supply trip distances, fares, and durations; the NGSIM vehicle trajectory dataset contributes lane-level vehicle speeds; Open-Meteo archives provide hourly meteorological variables for Los Angeles across 2023; and Chicago Traffic Tracker speed estimates feed a tertiary speed channel. Together these sources contribute 21,666 observations. However, the authors are explicit that the multi-modal samples are artificially assembled: each modality channel is independently sourced and combined under a generative process, not co-registered records of a single real logistics event. The delivery-delay targets themselves are computed by a non-linear interaction model with congestion, telemetry, and weather terms, clipped to a range of 5.5 to 18.5 minutes.

On the held-out test set of 1,500 samples, MARLIN achieved a mean absolute error of 0.7966 minutes, a root-mean-square error of 1.1014 minutes, an R-squared of 0.5856, a mean absolute percentage error of 7.88 percent, and 91.8 percent accuracy within a two-minute tolerance. The aggregate SLA compliance rate was 99.8 percent, meaning predicted delays stayed below the 15-minute service threshold in nearly all cases. The authors stress, however, that because the labels follow a known generative rule, these numbers quantify internal performance within the benchmark and cannot be read as validation on observed dispatch-to-delivery outcomes.

The architecture works in stages. Five encoders are first pre-trained independently with denoising autoencoder objectives: a bidirectional LSTM for the 24-step traffic sequence, a PointNet-style network for LiDAR features, a lightweight Vision Transformer for camera embeddings, and multilayer perceptrons for the eight-channel telemetry and six-variable weather inputs. The authors chose reconstruction over contrastive pre-training because contrastive losses degenerate at the small batch sizes typical of per-modality IoT datasets; an earlier contrastive variant of their architecture produced losses near the theoretical floor and a traffic-encoder R-squared of only about 0.05, whereas the reconstruction losses converged to between 0.006 and 0.023 across modalities.

After pre-training, the five 128-dimensional embeddings are fused through a two-layer stacked multi-head cross-attention module with learnable per-modality gates that can suppress degraded sensor streams. A dual-head prediction network then outputs both a delay estimate and an SLA-compliance logit, trained jointly with a Huber regression loss and a weighted binary cross-entropy loss, regularized with Mixup, exponential-moving-average weight averaging, and OneCycleLR scheduling. A three-layer GraphSAGE encoder embeds a 120-node, 400-edge logistics graph, and a Soft Actor-Critic agent with prioritized replay uses the fused state representation to select continuous routing actions over 1,500 training episodes.

The routing results are the most qualified in the paper. In the controlled graph environment, classical planners outperformed all reinforcement-learning variants: Dijkstra with free-flow edge costs achieved a mean delay of 7.80 minutes and A* with congestion-weighted costs 7.63 minutes, compared with roughly 8.7 to 9.3 minutes for the learning agents. The GraphSAGE-augmented SAC agent did not significantly outperform a flat-state SAC ablation at the tested seed count. The authors conclude that MARLIN’s routing contribution lies in integrating learned multi-modal state representations with sequential decision-making, not in beating classical shortest-path algorithms under fully observed static conditions, and they suggest the framework’s advantage would more plausibly appear in dynamic or partially observed networks.

The prediction comparison also requires nuance. Random Forest and Gradient Boosting, fed a 58-dimensional feature vector of temporal means and sensor summaries, achieved the best point-prediction accuracy on the benchmark, with mean absolute errors of 0.6937 and 0.6943 minutes respectively. MARLIN beat Ridge Regression, Bayesian Ridge, SVR, and KNN, improving on Ridge by 4.6 percent, but the authors state plainly that the framework should not be positioned as the most accurate standalone predictor. Its value, they argue, lies in processing raw sensor streams directly, providing cross-modal interpretability, SLA-aware multi-task learning, and routing functionality that tree ensembles lack.

Ablation studies produced an instructive surprise. A reduced configuration concatenating only Traffic and IoT embeddings without attention achieved the lowest mean absolute error of all variants, 0.7148 minutes, significantly better than the full model in a Welch’s t-test, though the effect was small. The authors attribute this to the synthetic delay function being dominated by traffic and IoT interaction terms, so the remaining modalities add mostly weak marginal signals. A leave-one-out analysis confirmed the hierarchy: removing Traffic raised the error by 39.9 percent, IoT by 25.0 percent, LiDAR by 11.9 percent, Camera by 9.9 percent, and Weather by 7.9 percent, an ordering corroborated by mutual-information and Random Forest feature-importance analyses.

The headline SLA figure deserves particular caution. Because only 4 of 1,500 test samples were true violations of the 15-minute threshold, a violation rate of roughly 0.27 percent, the aggregate compliance rate is dominated by easy cases. The confusion matrix shows MARLIN detected only 2 of 4 true violations, with 1 false alarm, yielding 50 percent recall and an F1 score of 57.1 percent for the violation class. The authors warn against reading the 99.8 percent compliance figure as evidence of reliable violation detection and recommend class-weighted losses or decision-threshold calibration in future work.

Graph-scaling experiments separated computational scalability from policy scalability. Message-passing cost grew approximately linearly across networks of 120, 240, and 480 nodes, but routing performance with a fixed training budget collapsed at the largest scale, with the destination-reach rate falling to 26.7 percent, and zero-shot transfer to larger networks remained inconsistent across seeds. All experiments ran on a single workstation with an 8 GB laptop GPU. The authors position MARLIN as a controlled architectural benchmark rather than evidence of operational superiority, and they identify validation on co-registered logistics data with observed dispatch-to-delivery outcomes, high-fidelity traffic simulators, and dynamic graph attributes as the essential next steps before any deployment claims can be substantiated.

Subject of Research: A multi-modal deep reinforcement learning framework for delivery-delay prediction and adaptive routing in IoT logistics, evaluated on a controlled benchmark with synthetic delay labels.

Article Title: Multi-modal deep reinforcement learning for traffic-aware adaptive routing in IoT logistics systems

Article References: Zhang, M. (2026). Multi-modal deep reinforcement learning for traffic-aware adaptive routing in IoT logistics systems. Discover Internet of Things, 6(1), Article 242. https://doi.org/10.1007/s43926-026-00534-0

Image Credits: AI Generated

DOI: 10.1007/s43926-026-00534-0

Keywords: IoT logistics, deep reinforcement learning, multi-modal fusion, traffic prediction, delivery delay, Soft Actor-Critic, GraphSAGE, cross-attention, SLA compliance, controlled benchmark, synthetic labels, adaptive routing