Scalable Attention Mechanisms for Real-Time Sensor Fusion in Autonomous Industrial Robots

We present SARTI, a transformer-based architecture fusing heterogeneous sensor streams — LiDAR, tactile, and infrared — at sub-10ms latency. Validated across seven industrial environments, achieving 97.3% object classification accuracy under heavy occlusion — a 14-point improvement over all existing baselines. Cross-modal positional encoding is the dominant contributor to performance gains.

ACCURACY — 7 ENVIRONMENTS (%)

#SensorFusion #Transformers #Robotics #RealTimeOPEN ACCESS

DS·B

IJMD Data Science@ijmd.data·5h agoSeries B · Vol 12

Federated Learning Under Non-IID Conditions: Convergence Analysis with Differential Privacy Guarantees

First tight convergence bounds for federated optimization under (ε,δ)-differential privacy with heterogeneous data. Reveals a fundamental tension between privacy budget and convergence rate — enabling practitioners to make principled trade-offs. Confirmed across four real-world healthcare and finance datasets within 2% margin.

#FederatedLearning #DifferentialPrivacy #ConvergenceOPEN ACCESS

FM·F

📌 SPECIAL ISSUE · CALL FOR PAPERS

IJMD Foundation Models@ijmd.fm·May 28Series F · Vol 12·3

Foundation Models in Industrial Systems — Now Accepting Submissions

Research on deployment, adaptation, and safety evaluation of large-scale pre-trained models in manufacturing, logistics, energy, and critical infrastructure. Deadline: Aug 15, 2026. IF 2.8 · Avg decision 18 days · Scopus & SCI-Expanded indexed.

#FoundationModels #IndustrialAI #LLMSafety #CFP

RB·C

IJMD Robotics@ijmd.robotics·8h agoSeries C · Vol 12

A Unified Framework for Sim-to-Real Transfer in Contact-Rich Manipulation Tasks

Addresses the sim-to-real gap for dexterous manipulation combining domain randomization, tactile simulation, and online adaptation. Transfers policies trained in simulation to seven real-world tasks with zero real-world data collection. Success rates 82–94% — state-of-the-art for contact-rich manipulation.

#SimToReal #DexterousManipulation #TactileSensing

EC·D

IJMD Edge Computing@ijmd.edge·12h agoSeries D · Vol 12

Neuromorphic Chips for Low-Power Edge Inference: Benchmarking Against GPU Baselines Across Eight Workloads

Comprehensive benchmark of Intel Loihi 2, BrainScaleS-2, and Akida against A100 and Jetson Orin across eight inference workloads. Neuromorphic hardware achieves 40–120× energy reduction with 3–8% accuracy degradation — a Pareto-optimal regime unavailable to conventional GPUs for always-on sensing.

#Neuromorphic #TinyML #EdgeAIOPEN ACCESS