AgentSkillsCN

Android Ml Ondevice

Android 设备端机器学习

SKILL.md

name: android-ml-ondevice description: On-device ML — ML Kit, TensorFlow Lite, Gemini Nano, model management. user-invocable: true

Android On-Device ML

Cross-cutting skill — always activates alongside the project-type-specific skill.

Hard Rules

  • •Prefer ML Kit for common tasks (text recognition, barcode, face, pose) — pre-trained, optimized.
  • •Run inference off the main thread — use coroutines with Dispatchers.Default.
  • •Use Google Play Services-based ML Kit models to reduce APK size.
  • •Bundle TFLite models in assets/ or download dynamically via Play Asset Delivery.
  • •Validate model input/output shapes — mismatches cause silent failures or crashes.
  • •Handle model loading failures gracefully — provide fallback or skip ML feature.
  • •Do not ship unnecessarily large models — quantize (INT8) to reduce size and latency.
  • •Respect user privacy — process data on-device, do not upload without consent.
  • •Request camera/microphone permissions before ML features that use them.
  • •Test ML features on low-end devices — not just flagships.
  • •Prefer MediaPipe Tasks API for new vision/text/audio ML features — unified, cross-platform, actively maintained.
  • •TensorFlow Lite is now rebranded as LiteRT — update references in new code.

Core Patterns

  • •ML Kit auto-downloads models via Google Play Services (no APK size impact).
  • •Use InputImage.fromMediaImage() for CameraX integration with ML Kit.
  • •TFLite Interpreter with GPU delegate for acceleration on supported devices.
  • •Use Gemini Nano (on-device LLM) via AI Core API for generative tasks — supports text and multimodal input.
  • •Implement progressive enhancement — ML features enhance but are not required.
  • •Cache inference results when input hasn't changed.

References

  • •references/ml_kit.md
  • •references/tensorflow_lite.md
  • •references/gemini_nano.md
  • •references/model_management.md
  • •references/mediapipe.md