Motion inference on Wear OS

This tutorial runs an Edge Impulse motion classification model on a Wear OS smartwatch. The app reads the accelerometer through SensorManager, fills a buffer sized to the model’s window, and runs the C++ classifier through JNI each time the buffer is full. Jetpack Compose for Wear OS shows the result on the watch face. The keyword spotting example uses the same buffered approach for audio on a phone.

Round Wear OS watch face listing motion classification labels and confidence scores — Wear OS app showing motion classification results on a round watch face

Prerequisites

A motion classification model trained in Edge Impulse on accelerometer data, and optionally gyroscope data
Android Studio with NDK 27.0.12077973 and CMake 3.22.1 installed
A Wear OS device or emulator running API 30 or later

1. Clone the repository

git clone https://github.com/edgeimpulse/example-android-inferencing.git
cd example-android-inferencing/example_motion_WearOS

2. Download TensorFlow Lite libraries

cd app/src/main/cpp/tensorflow-lite

# macOS/Linux
sh download_tflite_libs.sh

# Windows
download_tflite_libs.bat

3. Export your model

In Edge Impulse Studio, go to Deployment.
Select Android (C++ library).
Enable EON Compiler to reduce inference time and battery use.
Click Build and download the .zip file.

4. Integrate the model

Extract the downloaded .zip file and copy its contents, except CMakeLists.txt, into app/src/main/cpp/. The folder then looks like this:

app/src/main/cpp/
├── edge-impulse-sdk/
├── model-parameters/
├── tflite-model/
├── tensorflow-lite/
├── wearosinference.cpp
└── CMakeLists.txt   # keep the existing file

5. Match the input frame size

The model expects a fixed number of float values as input. Find the value in model_metadata.h:

#define EI_CLASSIFIER_DSP_INPUT_FRAME_SIZE 375

Then update MainActivity.kt to match:

// If your model uses 375 floats (125 samples × 3 axes), set this to 375
private const val EI_CLASSIFIER_DSP_INPUT_FRAME_SIZE = 375

If the values don’t match, the classifier returns incorrect results or crashes.

6. Build and run

Open the project in Android Studio.
Select Build -> Make Project.
Connect your Wear OS device or start a Wear OS emulator.
Run the app.

Perform one of your trained movements. The classification result appears on the watch face.

How it works

Sensor setup

The app registers the accelerometer listener in onResume and removes it in onPause, so the sensor doesn’t drain the battery while the app isn’t visible:

override fun onResume() {
    super.onResume()
    accelerometer?.also {
        sensorManager.registerListener(this, it, SensorManager.SENSOR_DELAY_NORMAL)
    }
}

override fun onPause() {
    super.onPause()
    sensorManager.unregisterListener(this)
}

Ring buffer collection

Each sensor event writes its x, y, and z values into the buffer. When the buffer is full, the app runs inference and resets the index:

private val ringBuffer = FloatArray(EI_CLASSIFIER_DSP_INPUT_FRAME_SIZE)
private var ringBufferIndex = 0

override fun onSensorChanged(event: SensorEvent) {
    when (event.sensor.type) {
        Sensor.TYPE_ACCELEROMETER -> {
            ringBuffer[ringBufferIndex++] = event.values[0] // x
            ringBuffer[ringBufferIndex++] = event.values[1] // y
            ringBuffer[ringBufferIndex++] = event.values[2] // z

            if (ringBufferIndex >= EI_CLASSIFIER_DSP_INPUT_FRAME_SIZE) {
                ringBufferIndex = 0
                runInference(ringBuffer)
            }
        }
    }
}

Native inference

The float array crosses the JNI boundary, and the native code passes it to the classifier:

extern "C"
JNIEXPORT jstring JNICALL
Java_com_edgeimpulse_edgeimpulsewearos_presentation_MainActivity_runInference(
        JNIEnv* env, jobject,
        jfloatArray data) {

    jsize len = env->GetArrayLength(data);
    jfloat* elements = env->GetFloatArrayElements(data, nullptr);
    std::vector<float> features(elements, elements + len);

    signal_t signal;
    signal.total_length = features.size();
    signal.get_data = &get_signal_data;

    ei_impulse_result_t result = {0};
    run_classifier(&signal, &result, false);

    return formatResult(env, result);
}

Wear OS UI

The app shows the result in a Jetpack Compose for Wear OS Chip:

@Composable
fun InferenceChip(inferenceText: String) {
    Chip(
        onClick = { },
        label = {
            Text(
                text = inferenceText,
                style = MaterialTheme.typography.body1
            )
        },
        modifier = Modifier.fillMaxWidth()
    )
}

Adding gyroscope data

If your model was trained with both accelerometer and gyroscope, register both sensors and write all six values per sample:

override fun onSensorChanged(event: SensorEvent) {
    when (event.sensor.type) {
        Sensor.TYPE_ACCELEROMETER -> {
            ringBuffer[ringBufferIndex++] = event.values[0]
            ringBuffer[ringBufferIndex++] = event.values[1]
            ringBuffer[ringBufferIndex++] = event.values[2]
        }
        Sensor.TYPE_GYROSCOPE -> {
            ringBuffer[ringBufferIndex++] = event.values[0]
            ringBuffer[ringBufferIndex++] = event.values[1]
            ringBuffer[ringBufferIndex++] = event.values[2]
        }
    }

    if (ringBufferIndex >= EI_CLASSIFIER_DSP_INPUT_FRAME_SIZE) {
        ringBufferIndex = 0
        runInference(ringBuffer.clone())
    }
}

The axis order in the buffer must match the order you used when you collected the training data.

Troubleshooting

Incorrect or unstable classifications

Check that EI_CLASSIFIER_DSP_INPUT_FRAME_SIZE matches your model’s input size.
Check that the sensor sampling rate matches your training data. If you trained at a higher rate, change SENSOR_DELAY_NORMAL to SENSOR_DELAY_GAME (about 50 Hz).
Check that the axis order in the buffer matches your training data.

High battery drain

Keep the unregisterListener call in onPause().
Lower the sampling rate to the minimum your model needs.
Enable EON Compiler when you build the deployment.

No sensor data on emulator

Use the emulator’s extended controls to inject sensor data.
Test on a physical Wear OS device for realistic sensor data.

OVERVIEW

END-TO-END

TOPICS

TOOLS

INTEGRATIONS

Prerequisites

1. Clone the repository

2. Download TensorFlow Lite libraries

3. Export your model

4. Integrate the model

5. Match the input frame size

6. Build and run

How it works

Sensor setup

Ring buffer collection

Native inference

Wear OS UI

Adding gyroscope data

Troubleshooting

Next steps

Resources

​Prerequisites

​1. Clone the repository

​2. Download TensorFlow Lite libraries

​3. Export your model

​4. Integrate the model

​5. Match the input frame size

​6. Build and run

​How it works

​Sensor setup

​Ring buffer collection

​Native inference

​Wear OS UI

​Adding gyroscope data

​Troubleshooting

​Next steps

​Resources

Prerequisites

1. Clone the repository

2. Download TensorFlow Lite libraries

3. Export your model

4. Integrate the model

5. Match the input frame size

6. Build and run

How it works

Sensor setup

Ring buffer collection

Native inference

Wear OS UI

Adding gyroscope data

Troubleshooting

Next steps

Resources