🎓 Lesson 3
D2
Sensor Fusion Fundamentals: GNSS, LiDAR, IMU, and UWB
Sensor fusion is like combining GPS, laser scans, motion sensors, and ultra-precise radio signals to give a self-driving mining truck the best possible understanding of where it is and how it’s moving — even when one sensor fails or gets blocked.
🎯 Learning Objectives
- ✓ Explain how GNSS, LiDAR, IMU, and UWB complement each other in AHS localization under mine-specific challenges
- ✓ Design a basic Kalman filter architecture for fusing GNSS position and IMU acceleration measurements
- ✓ Analyze sensor error budgets to identify dominant sources of positioning uncertainty in an open-pit environment
- ✓ Apply covariance propagation principles to quantify fused pose uncertainty at 95% confidence level
- ✓ Evaluate the impact of UWB anchor placement geometry on ranging accuracy using geometric dilution of precision (GDOP)
📖 Why This Matters
In autonomous haul trucks operating in deep, steep-sided open-pit mines, no single sensor works reliably all the time: GNSS signals vanish near highwalls; LiDAR performance degrades in blast-induced dust and rain; IMUs drift within seconds without correction; UWB requires precise anchor installation and clear line-of-sight. Sensor fusion isn’t optional—it’s the engineering safeguard that keeps AHS operational, compliant with ISO 26262-derived functional safety targets (ASIL-B), and capable of <0.3 m lateral positioning accuracy required for collision avoidance and precise dumping. Failure to fuse intelligently has led to real-world incidents—including misaligned dump positioning causing spillage, rehandling, and unplanned maintenance downtime.
📘 Core Principles
Sensor fusion rests on three pillars: (1) Sensor modeling—characterizing each sensor’s noise profile (e.g., GNSS white noise + multipath bias, IMU bias instability, LiDAR angular quantization error); (2) State representation—defining the fused state vector (typically [x, y, z, roll, pitch, yaw, vx, vy, vz, bx_gyro, by_gyro, bz_gyro, bx_acc, by_acc, bz_acc]); and (3) Estimation framework—choosing between linear (Extended Kalman Filter) or nonlinear (UKF, PF) methods based on system dynamics and observability. In mining, the Extended Kalman Filter (EKF) remains industry standard due to computational efficiency and proven robustness when augmented with zero-velocity updates (ZUPT) during truck idling and map-aided LiDAR feature matching against pre-surveyed pit wall geometry. Crucially, fusion must be *fault-aware*: real-time consistency checks (e.g., innovation-based chi-square tests) detect and isolate faulty sensors before they corrupt the state estimate—a requirement codified in SAE J3016 Level 4 operational design domain (ODD) validation protocols.
📐 EKF Position Update Covariance
The EKF update step computes the posterior state covariance matrix P⁺ after incorporating a GNSS measurement. This quantifies uncertainty reduction and guides trust allocation across sensors. The formula reflects how measurement noise R and observation Jacobian H shape confidence in the fused result.
💡 Worked Example
Problem: A Komatsu 930E AHS truck receives a GNSS pseudorange measurement. Pre-update state covariance diagonal elements for x/y position are [0.81, 0.81] m². Observation Jacobian H = [1 0], measurement noise covariance R = 2.25 m² (95% HDOP × 1.5 m typical GNSS error). Compute updated x-position covariance.
1.
Step 1: Compute Kalman gain K = P⁻ Hᵀ (H P⁻ Hᵀ + R)⁻¹ → K = 0.81 × 1 / (1×0.81×1 + 2.25) = 0.81 / 3.06 ≈ 0.265
2.
Step 2: Apply update: P⁺ = (I − K H) P⁻ → P⁺_xx = (1 − 0.265×1) × 0.81 = 0.735 × 0.81 ≈ 0.595 m²
3.
Step 3: Verify: Updated covariance (0.595 m²) corresponds to ~0.77 m 1σ uncertainty — a 24% improvement over prior 0.9 m 1σ estimate, consistent with GNSS aiding in open-sky conditions.
Answer:
The updated x-position covariance is 0.595 m², yielding a 1σ uncertainty of 0.77 m — well within the AHS target of ≤0.3 m 2σ (0.15 m 1σ) when combined with LiDAR/IMU.
🏗️ Real-World Application
At Rio Tinto’s Gudai-Darri mine (Western Australia), the AHS fleet uses tightly coupled GNSS-IMU-LiDAR-UWB fusion running on NVIDIA DRIVE AGX Orin. During a 2023 blast event, GNSS was degraded for 17 seconds due to ionospheric disturbance and highwall shadowing. The EKF seamlessly down-weighted GNSS inputs, relying on pre-mapped LiDAR terrain features (matched to a 5-cm-resolution digital elevation model), ZUPT-corrected IMU dead reckoning, and UWB ranging from 12 strategically placed anchors along haul roads. Position error remained <0.22 m (2σ), enabling uninterrupted haul cycle completion—demonstrating fusion resilience validated per ISO/PAS 21448 (SOTIF) requirements for hazardous scenario recovery.