Robot Localisation Using UHF-RFID Tags: A Kalman Smoother Approach

Farhad Shamsfakhr, Andrea Motroni, Luigi Palopoli, Alice Buffi, Paolo Nepa, Daniele Fontanelli

TL;DR

This paper presents a novel approach to robot localization in industrial environments using UHF-RFID tags and an Extended Kalman Smoother (EKF-RTS). By fusing RFID data with odometry, the method achieves centimeter-level accuracy and demonstrates robustness to missing measurements, making it a practical solution for autonomous navigation in complex settings.

What Problem Does This Paper Address?

The increasing adoption of Industry 4.0 technologies, particularly autonomous vehicles in smart warehouses and factories, necessitates reliable and affordable indoor localization solutions. Existing methods often struggle with accuracy, cost, or robustness in challenging industrial environments. While UHF-RFID technology offers a low-cost and versatile option for localization, challenges remain in achieving precise positioning, especially when dealing with signal ambiguities and intermittent measurements. This paper aims to address these limitations by developing a more robust and accurate localization system that can reconstruct the entire trajectory of a mobile robot.

Methods

The core of this work is the development of an Extended Kalman Smoother (EKF-RTS) that leverages UHF-RFID tag phase measurements and odometry data. The system models the robot's motion using a unicycle model and incorporates a signal model that accounts for the phase ambiguity inherent in RFID readings. The EKF-RTS is designed to process these measurements and provide an estimate of the robot's state (position and orientation) over time.

Key methodological contributions include:

  • EKF-RTS Smoother: Unlike standard Kalman filters that only use past measurements, the RTS smoother incorporates a backward pass to refine estimates using future information, leading to improved trajectory accuracy.
  • Phase-Based RFID Measurement: The paper focuses on using the phase of the RFID signal, which is more robust than RSSI (Received Signal Strength Indicator) in indoor environments, despite the challenge of phase ambiguity.
  • Handling Phase Ambiguity: The method addresses phase ambiguity by using the difference in phase measurements between consecutive time steps, effectively creating a relative measurement that is less sensitive to absolute phase offsets.
  • Robustness to Missing Measurements: A significant contribution is the demonstration of the smoother's ability to maintain accuracy even when RFID measurements are intermittently unavailable, a common issue in real-world deployments.
  • Fixed-Lag Smoother: The paper also explores a fixed-lag version of the smoother, which balances accuracy with computational efficiency for near real-time applications.

Datasets

The paper utilizes two main sources of data:

  • Simulated Data: Extensive Monte Carlo simulations were performed using a simulated warehouse environment in MATLAB. These simulations involved generating various robot trajectories with different velocities and random deployments of RFID tags to evaluate the algorithm's performance under controlled conditions.
  • Experimental Data: An actual indoor office-like environment was used for experimental validation. This environment was instrumented with UHF-RFID tags, and a custom-built unicycle mobile robot equipped with an RFID reader, LIDAR, and encoders was used. Ground truth trajectories were obtained using a state-of-the-art SLAM algorithm with a LIDAR sensor.

Key Insights and Ideas

  • Smoother is Better: The RTS smoother significantly outperforms a standard EKF in terms of trajectory estimation accuracy, especially in complex scenarios. This highlights the benefit of incorporating future information for refining past estimates.
  • Phase is Powerful (but tricky): While phase-based RFID measurements are more robust than RSSI, the inherent phase ambiguity requires careful handling. The proposed differencing approach effectively mitigates this challenge.
  • Missing Data is Not a Dealbreaker: The EKF-RTS demonstrates remarkable robustness to missing RFID measurements, a critical factor for real-world industrial applications where signal availability can be inconsistent. This is a key advantage over filter-based approaches that can degrade significantly with data gaps.
  • Window Size Matters (for real-time): For real-time applications, there's a trade-off between the smoother's window size (affecting accuracy) and computational cost. The paper identifies an optimal window size that balances these factors.
  • Robot Dynamics Influence Accuracy: The paper shows that lower forward velocities can lead to longer trajectories and potentially increased localization uncertainty, as the relative displacement between measurements becomes smaller, making them more susceptible to noise.

Key Results

  • Position Accuracy: The experimental analysis showed a position Root Mean Square Error (RMSE) of approximately 15 cm.
  • Orientation Accuracy: The orientation RMSE was found to be around 0.2 radians.
  • Robustness to Missing Data: Even with only 30-40% of available RFID measurements, the uncertainty growth remained acceptable, demonstrating high robustness.
  • Simulation Performance: Monte Carlo simulations with 10,000 trials showed that the SEKF (fixed-lag smoother) achieved lower RMSE for both position and orientation compared to the standard EKF.
  • Computational Trade-off: The study found that increasing the window size of the smoother generally reduced localization error but also increased computation time. An optimal window size of 55 samples was identified for near real-time execution.

Limitations and Future Work

  • Initial Condition Uncertainty: While the SEKF is shown to be robust to initial condition uncertainty, the paper notes that the slight increase in RMSE at the beginning of trajectories is due to a transient effect. Further investigation into mitigating this initial transient could be beneficial.
  • Tag Density: The paper suggests that a higher density of RFID tags or more RFID antennas could further improve localization accuracy. The current experiments used a limited number of tags.
  • Real-time Performance: While the fixed-lag smoother offers a path towards real-time operation, the full-length smoother is computationally demanding. Future work could focus on optimizing the smoother for even faster execution.
  • Environmental Factors: The paper acknowledges that multipath interference can increase the uncertainty of RFID measurements beyond the nominal thermal noise. Further research into advanced signal processing techniques to mitigate these effects could be valuable.
  • Complex Dynamics: The current work focuses on a unicycle robot model. Extending the approach to robots with more complex dynamics or in highly dynamic environments could be a future direction.

Summary updated

Abstract

Autonomous vehicles enable the development of smart warehouses and smart factories with an increased visibility, flexibility and efficiency. Thus, effective and affordable localisation methods for indoor vehicles are attracting interest to implement real-time applications. This paper presents an Extended Kalman Smoother design to both localise a mobile agent and reconstruct its entire trajectory through a sensor-fusion employing the UHF-RFID passive technology. Extensive simulations are carried out by considering the smoother optimal-window length and the effect of missing measurements from reference tags. Monte Carlo simulations are conducted for different vehicle trajectories and for different linear and angular velocities to evaluate the method accuracy. Then, an experimental analysis with a unicycle wheeled robot is performed in real indoor scenario, showing a position and orientation root mean square errors of 15 cm, and 0.2 rad, respectively.
BibTeX
@article{shamsfakhr2021robot,
  author = {Farhad Shamsfakhr and Andrea Motroni and Luigi Palopoli and Alice Buffi and Paolo Nepa and Daniele Fontanelli},
  title = {{Robot Localisation Using UHF-RFID Tags: A Kalman Smoother Approach}},
  journal = {Sensors},
  year = {2021},
  doi = {10.3390/s21030717},
}
Figure