RadarMind: cognitive AI with mmWave sensing for human-aware intelligent automation

Human-aware intelligent automation requires robust and privacy-preserving perception of human intent. Vision-based solutions can be sensitive to lighting and occlusion and may raise privacy concerns, motivating the use of alternative sensing modalities. This paper presents RadarMind, a radar-driven gesture-recognition framework that uses a single TI IWR1443 mmWave radar sensor. RadarMind employs a lightweight pipeline comprising radar-based data collection, signal preprocessing (clutter removal, alignment/normalization, and noise reduction), and compact feature extraction. RadarMind extracts a micro-Doppler map and a compact four-dimensional temporal-statistical signature (TSS) comprising motion energy entropy (MEE), energy rise time (ERT), temporal asymmetry index (TAI), and the zero-crossing rate of the energy derivative (ZCR-dE). The micro-Doppler map is encoded to generate FiLM conditioning parameters, while the TSS vector is encoded as a low-dimensional temporal representation. These two streams are fused in MindNet through micro-Doppler-conditioned FiLM modulation to classify dynamic hand gestures. Experimental evaluation on data collected from 20 participants across seven gesture classes demonstrates that RadarMind achieves competitive recognition performance while maintaining a lightweight model design and an interpretable temporal-statistical feature branch suitable for real-time, touch-less interfaces and smart-environment automation. The current evaluation is limited to isolated, single-user gestures in a fixed indoor setup; cross-room, cross-mounting, continuous-stream, and multi-person robustness remain future work.