<![CDATA[This research proposes a cross-domain autonomy framework for automation in biomedical and agricultural systems—spanning autonomous surgery, precision agriculture, and assistive exoskeletons—by integrating self-supervised multimodal perception, constraint-based planning–control (Model Predictive Control/MPC with safety filter Control Barrier Functions/CBF), and meta-learning-based fast personalization. The architecture is implemented on low-latency edge computing and evaluated through a standardized protocol linking technical metrics, safety, and operational benefits. In surgical tasks, the proposed method reduces path error by ~39% (1.7 mm vs. 2.8 mm), peak force by ~22% (3.2 N vs. 4.1 N), and constraint violations by ~93% (0.04 vs. 0.60 events/min), with latency <50 ms. In precision agriculture, the system improved weed detection mAP to 0.78, decreased geolocation error to 3.9 cm, reduced nozzle drift to 9.6 cm, and saved 34.5% on average input without sacrificing dose uniformity (CV 17.5% vs. 21.0%). In the exoskeleton, CBF-constrained few-shot personalization reduced metabolic cost by −12.4%, improved step symmetry to 0.90 and user comfort to 4.3/5, while reducing constraint violations by 10× (0.02 vs. 0.20 events/min). The key novelty lies in cross-domain transferable “autonomy primitives” with formal safety guarantees, combined with rapid adaptation at low data cost. Results demonstrate that the framework is safe, adaptive, and efficient, ready to accelerate translation from the laboratory to clinical practice, production fields, and rehabilitation.]]>
