Source-linked AI summary
IMU-Aided Correction of Orientation-Induced Ranging Error in Bluetooth Channel Sounding on Commercial Hardware
Mihir Bapat, Santosh Nagaraj
TL;DR
BLE Channel Sounding can estimate short-range distance, but the effect of device orientation on commercial-hardware ranging error had not been studied. This paper measures that effect and applies IMU-based Random Forest correction, achieving a 74.6% MAE reduction under Leave-One-Orientation-Out evaluation.
Problem
Published CS accuracy studies kept device orientation constant, leaving orientation-induced ranging error on commercial hardware unaddressed.
Method
The study collects CS measurements and six-axis IMU features across nine orientations and nine distances, then trains and evaluates a Random Forest correction model.
Results
74.6% MAE reduction was achieved by the Random Forest across all nine orientations under Leave-One-Orientation-Out evaluation.
Takeaways & Limitations
The onboard IMU on commercial hardware contains enough information to meaningfully correct orientation-related CS ranging error using machine learning.
Abstract
from arXiv · showhide
Bluetooth Low Energy Channel Sounding (BLE CS), standardized in Bluetooth Core Specification 6.0 (September 2024), enables distance estimation via phase-based ranging (PBR) and round-trip time (RTT). Although prior work has studied CS accuracy in configurations with a fixed orientation, no published work has studied how device orientation affects ranging error on commercial hardware. We present the first study of orientation-induced CS ranging error and a Machine Learning correction using IMU features. This study used the EFR32xG24 Channel Sounding Development Kit, the only commercial CS platform with an integrated six-axis Inertial Measurement Unit (IMU). Our results show that device orientation has a substantial effect on CS ranging accuracy; we found that a Random Forest model trained on the IMU derived orientation achieved a 74.6\% Mean Absolute Error (MAE) reduction under a Leave One Orientation Out Evaluation, demonstrating that IMU readings have potential to improve ranging accuracy.
I. INTRODUCTION
BLE Channel Sounding estimates short-range distance with standard BLE radios, but device orientation can affect accuracy through antenna-pattern-dependent multipath. This paper addresses the unstudied orientation effect on commercial hardware and evaluates IMU-based correction.
- BLE CS estimates sub-meter distances with standard BLE radios without complex antenna arrays.
- PBR uses accumulated phase differences across up to 72 channels, while RTT uses signal round-trip travel time.
- Antenna radiation patterns change significantly when devices rotate or tilt, affecting multipath propagation and potentially ranging accuracy.
- Published CS accuracy studies kept orientation constant, leaving orientation-based error compensation with IMU measurements as an open direction.
- The study characterizes nine orientations and nine distances, identifies roll–pitch asymmetry, and evaluates Random Forest correction using IMU features.The model is assessed with Leave-One-Orientation-Out evaluation.
II. BACKGROUND AND RELATED WORK
Prior work has examined BLE CS ranging, machine-learning correction, and antenna-orientation effects separately. However, the supplied literature identifies no peer-reviewed study isolating orientation-induced error in CS measurements.
- BLE CS research has covered pre-standardization phase-based ranging, PBR and RTT definitions, comparative evaluations, vehicle access, and indoor environments.
- Machine-learning approaches have reduced errors in CS, BLE multipath, and UWB ranging, but none addressed orientation-induced CS ranging error.Reported examples include RMSE reductions up to 0.4 m and UWB P90 error reduction from 58% to 15 cm on unseen trajectories.
- Earlier antenna-orientation studies found effects in UWB, RF space-diversity systems, and sensor-network multipath characteristics.
- No peer-reviewed study isolates the effects of device orientation on CS measurements.
III. METHODOLOGY
The methodology uses paired commercial EFR32xG24 boards to collect BLE CS measurements and onboard six-axis IMU readings, then combines them with ground-truth labels for Random Forest correction evaluation.
- Two commercial EFR32xG24 Channel Sounding Development Kits were used, with one reflector board and one customized initiator reading IMU measurements.
- The measurement pipeline synchronously captures CS distance estimates and six-axis IMU readings at 8.9 Hz, combines them with ground-truth labels, and trains a Random Forest model.
- The initiator ran Channel Sounding in PBR mode with dual-antenna polarization diversity on both boards.The RTL library produced filtered measurements at approximately 8.9 Hz.
- Each recorded entry included timestamp, CS distance, three accelerometer readings, and three gyroscope readings from the onboard IMU.The accelerometer and gyroscope values were recorded as raw ADC measurements with specified sensor ranges and scales.
B. Data Collection
Data collection tested BLE CS across nine distances and nine initiator orientations in an indoor corridor, producing 44,576 measurements. Accuracy was relatively strong at short range but degraded substantially beyond 5 ft.
- 44,576 measurements were recorded across 81 sessions spanning nine distances and nine orientations.Each configuration was recorded for 60 seconds at 8.9 Hz, yielding 520–1,185 samples; the reflector remained flat.
- The tested distances ranged from 1 ft (0.30 m) to 30 ft (9.14 m), while orientations included flat, roll, pitch, face-down, and upright configurations.
- 164.17 cm overall MAE and 207.24 cm RMSE were observed, primarily because error increased at longer distances.
- MAE ranged from 19.70 to 39.15 cm at 1–3 ft but increased to 195–266 cm beyond 5 ft.
B. Orientation Effect at Short Range
Device orientation substantially changes short-range CS ranging error, with roll left 90 producing especially high errors and large variation across orientations.
- At 2 ft, MAE varied nearly fivefold, from 9.3 cm for pitch toward 30 to 44.6 cm for roll left 90.The lowest short-range MAE was 9.29 cm, while the highest short-range MAE was 57.57 cm at 3 ft for roll left 90.
- Table I reports key per-orientation MAE values at short range, marking the best and worst orientation for each distance.
- At 1 ft, MAE was consistent across orientations, but roll left 90 showed high error and variance at 2 ft and continued high error at 3 ft.
- Rolling left 90 degrees at 18 and 25 ft produced the highest MAE among the tested orientation-distance combinations.
C. Roll vs. Pitch Asymmetry
Roll-axis orientations yield greater ranging error than pitch-axis orientations, while roll left 90 degrades further with distance beyond the general multipath effect.
- Roll-axis orientations had 165.0 cm combined MAE versus 140.9 cm for pitch-axis orientations, a 1.24 ratio.The difference is supported by the BRD2606A’s dual printed antennas being optimized for the horizontal plane.
- Rolling rotates both antennas away from intended polarization alignment, whereas pitching primarily changes elevation angle with less polarization impact.Pitch direction also matters: pitching toward the reflector consistently outperformed pitching away at short range.
- Beyond 12 to 18 ft, most orientations plateaued, but roll left 90 continued degrading in accuracy.The pattern suggests a severe antenna-alignment factor for roll left 90 beyond the general indoor multipath environment.
D. IMU Tilt Correlation
IMU-derived tilt angle is positively associated with CS ranging error, providing statistically significant orientation information for the correction model.
- The linear fit produced r = 0.156 for absolute error and r = 0.126 for signed error, both with p < 0.001.The analysis used N = 44,576 measurements.
- Larger IMU-derived tilt angles were associated with larger absolute ranging errors.
- Figure 3 plots ranging error against IMU-derived tilt angle with points colored by distance.
E. IMU-Based Correction
The paper uses IMU features to correct orientation-induced CS error and evaluates generalization with orientation- and distance-held-out validation.
- A random 80/20 split produced 96.9% MAE reduction but was considered unreliable because near-duplicate measurements could appear in training and testing.The study therefore reported two cross-validation strategies for a more honest generalization measure.
- Random Forest substantially outperformed Linear Regression and Multilayer Perceptron under LOOO without inflated accuracy.
- 74.6% MAE reduction was achieved by Random Forest under Leave-One-Orientation-Out cross-validation.The model was evaluated on completely unseen orientations, averaging performance across all nine held-out orientation tests.
- LOOO trains on eight orientations, tests on the ninth held-out orientation, and averages across all nine tests.
- Under Leave-One-Distance-Out evaluation, Random Forest achieved an average 16.1% MAE reduction across all distances.
V. DISCUSSION
Orientation affects CS ranging through antenna-pattern and multipath changes, while an IMU-informed Random Forest substantially reduces errors on unseen orientations and distances.
- Rolling or tilting the board directs its strongest antenna radiation away from the reflector, lowering SNR and increasing the relative contribution of multipath signals.
- 74.6% LOOO reduction shows that the Random Forest generalizes to unseen orientations.
- 16.1% LODO reduction shows that the model also applies to unseen distances.
VI. CONCLUSION
The study establishes that orientation significantly affects BLE CS accuracy and that onboard IMU measurements can support meaningful machine-learning correction on commercial hardware.
- The evaluation measured orientation effects using IMU measurements across nine orientations and nine distances.
- The results show that device orientation has a significant impact on CS ranging performance.
- The onboard IMU on commercial hardware contains enough information to meaningfully correct orientation-induced error using machine-learning approaches.
- Future work should test more environments and vary reflector orientation, which remained fixed in this study.