Toward Shape Memory Alloy Actuation in Exoskeletons: A Large Language Model-Assisted Fuzzy Control Architecture with a Servo-Based Proof of Concept

Shape-memory alloys (SMAs) offer low mass, compact integration, quiet operation, and high work density, but their hysteresis and coupled thermal–mechanical behavior complicate repeatable control. This study develops a modular architecture in which natural-language command interpretation is separated from motion-profile generation and from actuator-specific models and controllers. The implemented pilot path was evaluated on a five-servo robotic hand, while the common comparison layer defines interfaces for technology-specific numerical models reserved for the next implementation stage. A locally executed large language model (LLM) converted user commands into port:angle pairs, and the same semantic front-end was coupled to either a fixed-step or zero-order Sugeno motion-profile generator. The semantic layer was evaluated using 60 recorded commands: 27 supported commands and 33 commands requiring rejection. Exact target generation was obtained in 25 of 27 supported cases (92.6%), while 17 of 33 invalid or unsupported commands were correctly rejected (51.5%), giving a strict overall result of 42 of 60 cases (70.0%). Direct observation confirmed the expected physical outcome in 23 of 27 supported cases (85.2%). The mean local response time was 25.30 s, the median was 16.71 s, and the 95th percentile was 66.73 s, confirming that the LLM is suitable only for the supervisory semantic layer. Because the pilot setup did not use an external position sensor, the fuzzy module is interpreted as a command-profile generator rather than a closed-loop position controller. The results validate the layered interface and establish an experimental and computational basis for a subsequent quantitative comparison of SMA and servo actuation.