Sensor Bring-Up Guide¶

This guide walks through connecting and testing each sensor type incrementally, from one sensor to the full 12-sensor configuration.

Philosophy¶

Don't connect everything at once. Each sensor type has its own test class in test_sensor_subsets.py. Validate each type in isolation before combining. A failure in one sensor should not block testing another.

Phase 1: One IMU¶

1a. Software-only test (no hardware)¶

.venv/bin/python -m pytest tests/test_sensor_subsets.py::TestIMUOnly -v

This runs 4 tests with mocked IMU data and verifies the feature extractor produces valid features when only IMU data is present (EMG + FSR zeroed).

1b. Connect one BNO055¶

Flash the KB2040 with only chest IMU enabled: edit firmware/kb2040/src/main.cpp, comment out thigh_ok and thigh reads.
Flash the STM32 with all IMUs disabled except #0 (L wrist): edit firmware/stm32f405/src/imu_task.c, comment out IMUs #1–#3.
Both MCUs should show LED_BUILTIN on solid.
Connect the STM32 to PC, run serial monitor at 115200: one IMU should report ok.
Run .venv/bin/python -m host.serial_reader: frames with one non-zero IMU block.

1c. Verify feature extraction on live data¶

from host.serial_reader import SerialReader
from host.feature_extraction import FeatureExtractor

extractor = FeatureExtractor()
with SerialReader("/dev/ttyACM0") as reader:
    for frame in reader:
        feats = extractor.add_frame(frame)
        if feats is not None:
            # 1 IMU → 50 nonzero features, rest zero
            nonzero = (feats != 0).sum()
            print(f"nonzero features: {nonzero} (expected ~50 for 1 IMU)")
            break

Phase 2: Multiple IMUs¶

Re-enable the second IMU on each MCU (thigh on KB2040, R wrist on STM32).
Re-flash both MCUs.
Run pytest tests/test_sensor_subsets.py -k IMUOnly — tests 2 IMUs.
Live test: nonzero features should be ~100.
Enable remaining IMUs (#2 L arm, #3 R arm on STM32).
Verify 6 IMUs → ~300 nonzero features.

Phase 3: EMG¶

Connect the MyoWare forearm sensor to STM32 A0.
Verify ADC reading: serial monitor should show non-zero EMG values when the muscle is flexed.
Run pytest tests/test_sensor_subsets.py -k EMGOnly — 3 tests pass.
Connect the back sensor to A1, verify both channels.

Phase 4: FSR¶

Connect one FSR with 10 kΩ voltage divider to A2.
Press the FSR — serial monitor shows increasing value.
Run pytest tests/test_sensor_subsets.py -k FSROnly — 3 tests pass.
Connect remaining 3 FSRs to A3–A5, verify all 4 channels.

Phase 5: Two sensor types¶

.venv/bin/python -m pytest tests/test_sensor_subsets.py -k TwoTypes -v

Tests: IMU+EMG, IMU+FSR, EMG+FSR. Each verifies that the feature extractor correctly handles mixed sensor types where one type is zeroed.

Phase 6: All sensors¶

.venv/bin/python -m pytest tests/test_sensor_subsets.py -k AllSensors

With all 12 sensors connected and streaming:

from host.serial_reader import SerialReader
from host.feature_extraction import FeatureExtractor
from host.live_inference import LiveClassifier

classifier = LiveClassifier("models/worker_har_best.pt")
extractor = FeatureExtractor()

with SerialReader("/dev/ttyACM0") as reader:
    for frame in reader:
        feats = extractor.add_frame(frame)
        if feats is not None:
            label, conf = classifier.predict(feats)
            print(f"{label} ({conf:.2%})")

The classifier should produce non-trivial predictions. If all labels are standing with low confidence, check that IMU quaternions are changing (the BNO055 is in NDOF mode and has calibrated).

Phase 7: Full pipeline¶

Flash both MCUs (reset to full firmware — all IMUs, EMG, FSR enabled)
Run unit tests: pytest tests/ -q → 127 passed
Start dashboard: streamlit run host/dashboard.py
Verify stick figure moves, EMG bars respond to muscle flex, FSR to foot pressure
Dashboard activity gauge updates ~10× per second

Common bring-up issues¶

Symptom	Phase	Fix
IMU `fail` on serial log	1b	Check I²C wiring, pull-ups, ADR pin
All features zero	1c	BNO055 not in NDOF mode? Check fusion mode
EMG always 0	3	MyoWare gain too low? Adjust onboard potentiometer
FSR saturation (always 1.0)	4	Voltage divider resistor too large; use 10 kΩ
Classifier always `standing`	6	Model not trained on your data; train with your recordings
Dashboard FPS < 5	7	Reduce stick-figure samples, close other tabs