❓ Q&A 25.1 — SensorFusion

[web3guru888]

❓ Frequently Asked Questions — SensorFusion (Phase 25.1)

Reference spec: #562 | Architecture: #567

---

Q1: Why EKF over UKF or particle filter?

A: The Extended Kalman Filter provides an excellent balance of computational efficiency and accuracy for mildly nonlinear systems. For sensor fusion in embodied cognition:

• EKF — O(n²) per update, sufficient for state dimensions ≤ ~50, handles moderate nonlinearity via Jacobian linearisation
• UKF — O(n³) sigma-point propagation, better for highly nonlinear models but heavier
• Particle filter — O(k·n) with k particles, handles arbitrary distributions but expensive for high-dimensional state

The SensorFusionConfig.filter_type field allows switching between backends. For most embodied scenarios the EKF is the right default. The Protocol interface (SensorFusion.fuse()) is filter-agnostic so swapping is transparent.

---

Q2: How does multi-rate fusion handle different sensor frequencies?

A: Three mechanisms:

1. Downsampling — High-rate sensors (e.g., tactile at 1000 Hz) are buffered and resampled to the fusion tick rate (e.g., 100 Hz) using windowed averaging
2. Interpolation — Low-rate sensors (e.g., vision at 30 Hz) are interpolated (linear/cubic) to produce estimates at the fusion tick
3. Predict-only steps — Between actual measurements, the EKF runs pure prediction (state + covariance propagation), so the state estimate is always current

The temporal alignment layer timestamps every reading with monotonic nanosecond precision and aligns them to a common timeline before feeding the EKF.

---

Q3: What happens when a sensor goes offline?

A: Graceful degradation:

1. After config.timeout_ms with no readings, the sensor is marked STALE
2. Its contribution to the Kalman gain is zeroed (effectively: the observation matrix H row is removed)
3. The corresponding covariance block grows without bound (no measurement updates)
4. FusedState.confidence is scaled down proportionally to the fraction of active sensors
5. If all sensors are offline, the system enters DEAD_RECKONING mode (predict-only) and emits a Prometheus alert

Recovery is automatic — when readings resume, the sensor is re-registered and the next update incorporates its observation.

---

Q4: How is the Mahalanobis distance threshold chosen?

A: The threshold comes from the χ² distribution with degrees of freedom equal to the observation dimension:

d²_threshold = χ²_inv(1 - α, dim)

For example, with α = 0.01 (99% confidence) and dim = 3 (3D position):

• d²_threshold ≈ 11.345

Any reading with Mahalanobis distance above this is rejected as an outlier. The α parameter is configurable via SensorFusionConfig.gate_alpha.

This is a standard technique in target tracking (Bar-Shalom & Li 1995) — it rejects spurious readings while accepting the natural spread of sensor noise.

---

Q5: How does calibration work?

A: The calibrate(sensor_id, n_samples) method:

1. Collects n_samples raw readings from the specified sensor
2. Computes the empirical mean (bias) and covariance
3. Updates the sensor's noise model: R_calibrated = empirical_covariance
4. Stores the bias vector for online subtraction: z_corrected = z_raw - bias

Calibration should be run at startup or when sensor characteristics change (e.g., temperature drift). The NullSensorFusion implementation skips calibration entirely.

---

Q6: Can new sensor modalities be added at runtime?

A: Yes. The register_sensor(sensor_id, modality, config) method:

1. Extends the state vector with new dimensions if the modality introduces new state variables
2. Allocates the corresponding observation matrix H rows
3. Initialises the noise covariance R block from the sensor config
4. Starts accepting readings immediately

The SensorModality enum can be extended without breaking existing registrations. The LiveModuleOrchestrator (Phase 15) can trigger sensor registration/deregistration as modules are hot-swapped.

---

Q7: How does SensorFusion integrate with WorldModel?

A: The integration path:

SensorFusion.fuse() → FusedState
  → WorldModel.update_belief(fused_state)
    → Belief state updated with position, velocity, covariance
    → SurpriseDetector (13.4) checks for unexpected state changes
    → CuriosityModule (13.3) evaluates information gain

The FusedState.covariance matrix is particularly valuable for the WorldModel — it provides calibrated uncertainty that feeds into downstream decision-making (Phase 23 DecisionTreeSolver).

---

Have more questions? Comment below!