Polaris — Vision-free Magnetic Fiducials

📄 Paper (ACM MobiCom 2024) 🧑‍💻 Code of Polaris 🎥 Demo of Polaris

Polaris is a vision-free fiducial system built on a magnetic constellation of passive disc magnets and a lightweight magnetometer array.It provides reliable relative pose estimation and high-capacity ID encoding even under occlusion, dust, low light, and privacy constraints. The tags are battery-free and low-cost, making Polaris a practical complement to existing camera-based fiducials.

Motivation — a complementary path when vision degrades

Vision-based fiducials (e.g., AprilTag) are sensitive to visibility (e.g., occlusion, low light) and can raise power and privacy concerns. In contrast, magnetic sensing is robust to occlusion and penetrates common obstructions, providing a low-power, camera-free complementary signal that preserves robust relative pose and high-capacity IDs when vision degrades.

System at a glance

Polaris is a full-stack magnetic fiducial with two core components:
1) a compact constellation tag — MOSK-coded passive disc magnets where bits are carried by spatial placement and polarity (N/S) orientations;
2) a linear magnetometer array with a lightweight decoding pipeline that performs magnetic field sampling → magnet detection & polarity inference → constellation reconstruction, yielding ID decoding and relative pose estimation.

(A) shows the magnetic sensing array of Polaris on different robotic systems; (B) and (C) compare the legacy ground-mounted vision-based tags with Polaris tags. Note they have the same **encoding capacity** and posture calibration capability.

How Polaris works — design & innovations

Polaris realizes a vision-free fiducial by replacing visual patterns with a magnetic constellation that encodes information through the orientation of diametrically magnetized discs and decodes it via a linear magnetometer array.

1. MOSK-coded constellation

Each tag adopts Magnetic Orientation-shift Keying (MOSK), a digital modulation scheme where every magnet stores an M-ary symbol using its dipole orientation (e.g., eight discrete angles -> 3 bits). Unlike visual textures or printed codes, orientation is translation-invariant and insensitive to small lateral displacements, allowing dense placement within a compact footprint. A typical 3 × 3 cm tag with nine magnets achieves AprilTag-level payloads while remaining passive, thin, and durable. Optional spatial permutation and CRC/FEC layers enhance ID robustness against missing or flipped magnets.

2. Sensing and decoding pipeline

A linear magnetometer array samples tri-axial magnetic fields as the robot passes over the tag. Polaris employs a lightweight derivative-based detector to identify field zero-crossings corresponding to individual magnets and their polarities. For each detected peak, a short template matching procedure based on Dynamic Derivative Time Warping (DDTW) estimates the local dipole angle, reconstructing the constellation pattern. Using the known array geometry and motion odometry, the system infers each magnet’s position and the tag’s relative pose (x, y, heading), followed by symbol demodulation and CRC/FEC verification to output a stable ID + pose stream in real time on embedded hardware.

3. Hardware design and deployment

The sensing bar integrates low-cost Hall-effect magnetometers on a modular PCB, connected to an embedded controller (e.g., nRF52832). The architecture emphasizes low power, low latency, and ease of integration with mobile robots or sensor heads, enabling camera-free fiducial tracking even on resource-constrained platforms.

Experimental evaluation — platforms, setup, and key results

We extensively evaluated Polaris across two robotic platforms (robot car and mini car) and a range of operating conditions (speed, height, lateral offsets, debris/coverings, and illumination). Figure A summarizes the sensing module and testbeds; Figure B shows the end-to-end (E2E) pipeline in operation—ID + relative pose streamed in real time under degraded visibility—demonstrating the system’s usability.

Key results

• Pose & decoding accuracy. Mean Euclidean error 0.58 mm (STD 0.08 mm); mean heading error 0.997° (STD 0.125°). With 8-level MOSK, the embedded decoder on ESP32 achieves BER ≈ 0.033.
• Tiny tags & frugal power. Tags as small as 1.6 × 1.6 cm² remain decodable; a three-sensor bar runs at ≈ 25.08 mW, suitable for miniature or solar-powered robots.
• Robust usability. Reliable ID + pose under occlusion, dust, low light, and modest lateral misalignment, validating Polaris as a camera-free complement to visual fiducials.

Experiment setup on two robotic platforms.

End-to-end posture calibration of a robot car through dust occlusion.

Publication

Jike Wang, Yasha Iravantchi, Alanson Sample, Kang G. Shin, Xinbing Wang, and Dongyao Chen. 2024. Polaris: Accurate, Vision-free Fiducials for Mobile Robots with Magnetic Constellation. Proceedings of the 30th Annual International Conference on Mobile Computing and Networking (ACM MobiCom ’24). Association for Computing Machinery, New York, NY, USA, 1560–1574. DOI: 10.1145/3636534.3690711

@inproceedings{Polaris_MobiCom24,
author = {Wang, Jike and Iravantchi, Yasha and Sample, Alanson and Shin, Kang G. and Wang, Xinbing and Chen, Dongyao},
title = {Polaris: Accurate, Vision-free Fiducials for Mobile Robots with Magnetic Constellation},
year = {2024},
isbn = {9798400704895},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3636534.3690711},
doi = {10.1145/3636534.3690711},
pages = {1560–1574},
numpages = {15},
keywords = {magnetic sensing, fiducial system, magnetometer},
location = {Washington D.C., DC, USA},
series = {ACM MobiCom '24}
}