The global rise in the aging population presents growing challenges for maintaining physical activity, rehabilitation, and independent living. While digital exercise tools have emerged, many fail to balance performance, privacy, and accessibility—particularly for older adults using low-power devices. This paper introduces ARISE, a multi-modal, privacy-preserving AI system that integrates human pose estimation and conversational interaction to create an engaging at-home exercise experience. Implemented entirely on the network edge, ARISE leverages the Raspberry Pi 5 and Hailo-8 Neural Processing Unit (NPU) to achieve real-time, on-device computation without reliance on cloud services. The system combines a YOLOv8m-Pose model for body tracking with a voice-based conversational pipeline using Vosk (speech-to-text), LLaMA (language model), and Kokoro-TTS (speech synthesis) for natural feedback and guidance. Experimental results demonstrate significant reductions in inference latency with NPU acceleration and optimized conversational responsiveness through adaptive chunking and segmentation strategies. The project is open-sourced and available at https://github.com/BlueJayADAL/SCARP25_ARISE.

@inproceedings{paperbot11539,
  title = {ARISE: Real-Time Multi-Modal AI for Senior Wellness on Edge Devices},
  author = {Alexander Roop and Aaron Frist},
  booktitle = {ICSE},
  year = {2026},
  doi = {10.1145/3774748.3787752}
}