<p>With the convergence of large language models (LLMs) and wearable technologies, intelligent sports and health management based on natural language interaction has emerged as a frontier research direction. However, existing approaches in outdoor scenarios are often constrained by semantic ambiguity arising from unimodal perception, while generative recommendations lack closed-loop feedback grounded in real-time physiological states; moreover, data privacy concerns further hinder continuous model evolution. To address these challenges, this paper proposes the AIMHI-Sport (AI-Driven Mobile Human–Computer Interaction framework for Sports) framework. First, a vision–inertial multimodal projection mechanism is constructed to align environmental context with the semantic space of LLMs, thereby mitigating ambiguity in physical feature interpretation. Second, an online reinforcement learning agent is designed to dynamically refine LLM-generated planning recommendations based on real-time heart rate feedback. Finally, a FedLoRA architecture is introduced to enable collaborative evolution at the edge under extremely low communication overhead. Experimental results demonstrate that the proposed method achieves an activity recognition F1 score of 91.2% on the WEAR dataset, outperforming the baseline MotionBERT without vision-inertial semantic alignment (88.6%). In the PAMAP2 regulation task, the heart rate control error (RMSE) is reduced to 4.1&#xa0;bpm, while the time-in-target-zone (TiTZ) increases to 82.6%, representing improvements of 54% and 21%, respectively, over the static planning baseline MotionGPT. In addition, compared with the full-parameter federated approach FedALA, the communication overhead is reduced by 99.3%. This study explores a closed-loop computational framework from perception to regulation, providing a generalizable reference for the deployment of embodied intelligence in active health applications.</p>