Research / On-device AI
Research pillar 04

Interpreting smell without relying on the cloud.

Robots need fast local decisions. Smell AI has to run on small hardware, on limited power, in the field — interpreting changing chemical patterns where they happen.

What we research

Mobile platforms cannot always wait for cloud analysis. Aeralyte targets on-device inference on the same low-power hardware that does the sampling — small models, compressed, running against a live sensor stream.

Technical terms: edge AI, TinyML, neuromorphic AI, state-space models, model compression.

The pipeline

The same loop runs end to end: air is sampled, the sensor chamber responds, the response becomes a fingerprint, and an on-device model returns a label and a confidence — feeding a robot or IoT decision.

Input114 features from 5 sensor families
Targetsclean vs post-use · event family · time bucket
Outputsmell fingerprint + confidence

Findings

Our baseline model already clears the lab's model bar — clean-vs-post-use, IQOS-vs-vape, a rough time bucket, and a false-alarm threshold — on a scenario-heldout split. That fixes the data contracts and behavior the on-device model has to meet on real hardware.

114
input features
3
prediction targets
0.90
baseline acc
0.00
false alarm

Next: latency, memory, and power budgets, measured with the model running on the ESP32-S3 itself.

References

  • 01AERALYTE_RESEARCH_LAB_BRIEF.md — pillar 4, on-device AI.
  • 02Kimi_Agent · AI Pipeline Research Prompts — edge inference & model compression.
  • 03Kimi_Agent · Robot Olfactory AI Roadmap — TinyML / state-space directions.
  • 04GitHub — XoAnonXo/aeralyte.