WaveLink

Framing the Problem

In environments where reliability and safety matter most, traditional control systems fail.

• In sterile medical and lab settings, touching interfaces introduces contamination risk

• In industrial environments, physical controls can be unsafe or impractical

• For people with mobility impairments, touchscreens and voice assistants are often unreliable or unusable

• In smart homes, gesture systems are expensive ($300+) and cloud-dependent

Across all these cases, the same issue appears:
existing solutions are too expensive, too slow, or too dependent on infrastructure.

Idea Explanation

Wavelink is a low-cost, wearable gesture control system that runs a neural network entirely on-device.

• A wrist-mounted device detects hand motion

• A TinyML model classifies gestures in 29ms

• Commands are sent wirelessly to control real-world devices

• No cloud, no internet, no external dependencies

Instead of relying on expensive hubs or cloud APIs, Wavelink creates a self-contained system:

→ Gesture → On-device AI → Wireless command → Physical action

This makes it:

• Fast (sub-50ms response)

• Private (no data leaves the device)

• Reliable (works anywhere, even offline)

• Affordable ($23 total build cost)

Implementation

Wavelink is built as a two-node system: a wearable controller and a local actuator station.

1. Wearable (Input + Intelligence)

• Microcontroller: Raspberry Pi Pico 2 WH

• Sensor: MPU6050 (6-axis IMU @ 100Hz)

• Collects motion data (acceleration + rotation)

• Runs DSP + neural network inference directly on-chip

2. Signal Processing Pipeline

• Sliding window of motion data (2 seconds, 1,200 values)

• FFT extracts frequency-domain features

• Data compressed into 33 features

3. Machine Learning

• 2-layer neural network (20 → 10 neurons)

• ~1,500 parameters (6.2 KB)

• Outputs 5 gesture classes with confidence score

• Runs in ~1ms inference time

4. Communication

• Sends gesture via UDP over WiFi

• Arduino acts as a self-hosted access point

• No router or internet required

• <1ms transmission time

5. Demo Station (Output)

• Arduino Uno R4 WiFi receives commands

• Controls:

  • Servo arm

  • Relay (lamp)

  • Motor (bidirectional)

  • LCD + LEDs + buzzer

End-to-end system:
→ Wrist motion → AI classification → UDP packet → Physical action (<50ms total)

Challenges

1. Running AI on a microcontroller

• Limited RAM and compute

• Solved by:

  • Using feature extraction (FFT) instead of raw data

  • Designing a small, efficient neural network

2. Gesture ambiguity

• Similar gestures (e.g., flick up vs down) overlapped

• Solved by:

  • Using frequency-domain features

  • Designing gestures that activate different sensor axes

3. Real-time performance

• Needed instant response to feel natural

• Solved by:

  • On-device inference (no cloud latency)

  • UDP instead of TCP (no handshake overhead)

4. Hardware integration under time pressure

• First time using CAD for enclosure

• Completed 3 design iterations in 24 hours

• Achieved a compact, wearable form factor

Accomplishments

• Built a fully working AI-powered wearable system in 24 hours

• Achieved:

  • 29ms inference latency

  • <50ms total system response

  • 83.3% classification accuracy

  • Zero false triggers in idle state

• Designed and printed a custom enclosure from scratch

• Created a fully offline, infrastructure-independent system

Most importantly:
Proved that real-time AI can run on a $7 microcontroller without cloud support.

Next Steps

Wavelink has strong potential beyond a prototype.

Short-Term Improvements

• Increase dataset size → improve accuracy

• Add more gesture classes (5 → 15+)

• Improve flick gesture reliability

Hardware Enhancements

• Switch from WiFi → Bluetooth Low Energy (BLE) (10× lower power)

• Design a custom PCB for smaller form factor

• Reduce size to match a fitness band

Software & ML

• Add per-user fine-tuning (personalized models)

• Implement hierarchical gesture classification

• Improve robustness across users

System Expansion

• Multi-device control (one wearable → multiple outputs)

• Mesh network for collaborative environments (e.g., surgical teams)