On boot, before the main loop, sample 1000 gyro readings (~1s) while
board is held still. Compute per-axis mean offset (sensor-frame raw LSBs)
and subtract from all subsequent readings in mpu6000_read().
- mpu6000_calibrate(): LED1+LED2 solid ON during 1s sample window,
resets filter state to zero once bias is known
- mpu6000_is_calibrated(): gate; main loop blocks arming and USB
streaming until calibration completes
- Bias subtracted in sensor frame before CW270 axis transform + scale,
so all three axes (pitch/roll/yaw rate) benefit
- config.h: GYRO_CAL_SAMPLES=1000
- No flash storage — recalibrate fresh each boot (bias varies with temp)
Closes#21 (3.5°/s yaw drift), #23 (gyro bias calibration on boot).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The MAMBA F722S mounts MPU6000 at CW270 (clockwise 270°) which applies
rotation matrix R = [[0,1,0],[-1,0,0],[0,0,1]] to transform sensor axes
to board axes (Betaflight convention).
Firmware (mpu6000.c):
- accel_pitch: was atan2(ax, az) → now atan2(ay, az)
board_forward = sensor_Y, so ay drives pitch not ax
- accel_roll: was atan2(ay, az) → now atan2(-ax, az)
board_right = -sensor_X, so -ax drives roll not ay
- gyro_pitch_rate: was +raw.gx → now -raw.gx
board_gy (pitch) = -sensor_gx after R_CW270 transform
- gyro_roll_rate: raw.gy unchanged (board_gx = sensor_gy ✓)
- gyro_yaw_rate: raw.gz unchanged ✓
UI (index.html) rotation sign fixes:
- roll → -rotation.z: Three.js +z = CCW from camera = left bank;
our convention is right-bank-positive so negate
- yaw → -rotation.y: Three.js +y = CCW from above; sensor_Z points
down on MAMBA (az ≈ +1g when level) so gz+ = CW physical; negate
- pitch → +rotation.x: correct as-is (Three.js +x tilts nose up ✓)
Closes#15.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds motor_driver.c/h between the balance PID and the raw
hoverboard UART driver:
- Differential drive: balance_cmd → speed, steer_cmd → steer
- Steer-only ramping at MOTOR_STEER_RAMP_RATE (balance PID keeps
full immediate authority — no ramp on speed channel)
- Headroom clamp: reduces steer so |speed|+|steer|<=MOTOR_CMD_MAX
ensuring ESC never clips the balance command
- Emergency stop: latches on TILT_FAULT, clears on BALANCE_DISARMED;
send path stays in 50Hz ESC tick to avoid flooding UART
main.c: replace bare hoverboard_send() with motor_driver_update();
config.h: MOTOR_CMD_MAX=1000, MOTOR_STEER_RAMP_RATE=20 counts/ms
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Issue #12 — Roll displayed as pitch:
- Firmware was sending r=pitch_rate (wrong). Changed to r=roll_deg*10.
- mpu6000.c: add roll complementary filter (accel atan2(ay,az) +
gyro gy integration, same COMP_ALPHA=0.98 as pitch).
- IMUData: add roll and yaw fields.
- UI: updateIMU() now uses data.r/10 for roll (not client-side filter
that computed from ax/ay/az which firmware never sent).
- Three.js: roll -> rotation.z (banking), pitch -> rotation.x (tipping)
— axes were already correct, fix was the firmware data.
Issue #13 — Add yaw telemetry:
- Firmware: gyro Z integration (gz * dt) → s_yaw, sent as y=yaw_deg*10.
Gyro-only, drifts — no magnetometer.
- IMUData: yaw field added.
- UI: yaw -> rotation.y (spinning on vertical axis). Displayed in HUD.
- UI: YAW RESET button captures current yaw as new zero reference
(client-side offset, no new firmware command needed).
Closes#12, #13.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Root cause 1 (IWDG reset loop): safety_init() was called before
mpu6000_init() — IWDG 50ms timeout fires during ~510ms IMU init,
causing infinite MCU reset. Moved safety_init() to after all
peripheral inits (IMU, hoverboard, balance).
Root cause 2 (DCache coherency): USB TX/RX buffers merged into a
single 512B-aligned struct in usbd_cdc_if.c. MPU Region 0 configured
non-cacheable (TEX=1, C=0, B=0) in usbd_conf.c USBD_LL_Init() before
HAL_PCD_Init(). DCache stays ON globally — MPU handles coherency.
Removed SCB_DisableDCache() from main.c (caused boot crash).
Also: fix safety.c IWDG_RELOAD macro (float literals not valid in
#if); add crsf.c stub so crsf_state links (UART not yet wired).
Fixes issue #9.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Safety systems implementation:
IWDG Hardware Watchdog (50ms timeout, config.h WATCHDOG_TIMEOUT_MS):
- safety_init() configures IWDG at PSC/32 (0.8ms tick), reload=62
- safety_refresh() must be called every loop iteration
- Cannot be disabled once started — MCU resets if loop hangs
- Started after 3s USB init delay (avoids spurious startup reset)
Arm Hold Interlock (3s, config.h ARMING_HOLD_MS):
- Arm command starts a hold timer, not immediate motor enable
- Motors only enable after ARMING_HOLD_MS consecutive hold
- Disarm or tilt > 10° cancels pending arm
- Prevents accidental arm from single keypress
Tilt Fault Alert:
- safety_alert_tilt_fault() fires one-shot buzzer on TILT_FAULT edge
- Rider hears alarm when tilt cutoff triggers
- Edge-detected (buzzer only fires once per fault event)
RC Timeout (infrastructure):
- safety_rc_alive() checks crsf_state.last_rx_ms vs RC_TIMEOUT_MS
- RC disarm wired but guarded (no CRSF yet) — remove guard when wired
- Compatible with future CRSF implementation
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add USB command interface for live PID gain adjustment without reflashing:
P<kp> I<ki> D<kd> T<setpoint_deg> M<max_speed> ?
Command parsing runs in main loop (sscanf-safe), not in USB IRQ.
USB IRQ copies command to shared volatile buffer (cdc_cmd_buf), sets flag.
Acknowledgement echoes current gains: {"kp":...,"ki":...,"kd":...}
Bounds checking: kp 0-500, ki/kd 0-50, setpoint ±20°, max_speed 0-1000.
Gains validated before write — silently ignored if out of range.
Telemetry updated from raw counts to physical tuning signals:
pitch (°x10), pitch_rate (°/s x10), error (°x10),
integral (x10 for windup monitoring), motor_cmd, state
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Bug 1 (PRIMARY — DCache/USB coherency):
SCB_DisableDCache() was buried inside icm42688_init(), called ~3.5s
after USB starts. STM32F7 DCache/USB coherency issue: when DCache is
on (enabled by SystemInit()), CPU writes to TX buffers stay in cache
and the USB hardware reads stale SRAM data. Moved SCB_DisableDCache()
to main() before HAL_Init(), ensuring coherency for all USB transfers.
Bug 2 (USB TX corruption):
CDC_Transmit() passed the caller's stack-allocated buf pointer directly
to the USB stack. The USB TXFE interrupt fires asynchronously; by then
the stack buffer may have been modified by the next loop iteration.
CDC_Transmit() now copies into the static UserTxBuffer before handing
off to the USB hardware, ensuring the buffer is stable for the transfer.
Bug 3 (IMU type mismatch → wrong data to balance):
main.c called icm42688_init()/icm42688_read() directly, passing
icm42688_data_t* (raw int16 ax/ay/az/gx/gy/gz) to balance_update()
which expects IMUData* (float pitch/pitch_rate from complementary
filter). Type mismatch produced garbage balance values. Fixed by using
mpu6000_init()/mpu6000_read() which wraps icm42688 + sensor fusion.
Telemetry updated to report fused pitch/rate instead of raw ADC counts.
Also fix icm42688_init() returning 0 on who==0 (no SPI response),
which falsely indicated IMU success.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
