A Robust Localization, Slip Estimation, and Compensation System for WMR in the Indoor Environments

Zakir Ullah, Zhiwei Xu, Lei Zhang, Libo ZhangSymmetry 2018 · DOI

TL;DR

This paper tackles the challenge of accurately locating wheeled mobile robots (WMRs) in environments where wheel slip is a significant issue, which traditional pure-rolling assumptions fail to address. They propose a novel system that combines DecaWave ranging with advanced heading estimation techniques (Reinforcement Learning and Location-based Heading Estimation) to improve localization accuracy, and then use this information to estimate and compensate for lateral slip.

What Problem Does This Paper Address?

Accurate localization is crucial for autonomous robots, but many existing systems assume wheels roll purely on the ground. In reality, factors like slippery surfaces, high speeds, and sensor noise can cause wheels to slip, leading to significant deviations from the robot's intended path and inaccurate position estimates. This problem is particularly challenging in indoor environments due to factors like multipath fading and non-line-of-sight conditions. Existing localization methods, especially those relying on heading information, are highly sensitive to even small errors in angular estimation, which are exacerbated by wheel slip. This paper aims to bridge the gap between planning and navigation by developing a robust system that can handle unknown lateral slip and improve WMR localization and tracking performance in such challenging conditions.

Methods

The core of the proposed system involves several key components:

  • DecaWave-based Real-Time Location System (RTLS): This system uses Ultra-Wideband (UWB) technology to provide noisy location estimates of the WMR. Trilateration is employed to calculate probable locations, and these measurements are also used to update Particle Filter (PF) weights.
  • Particle Filter (PF): A standard Bayesian filtering technique used for state estimation, particularly effective in handling non-Gaussian uncertainties common in environments with multipath fading and NLOS conditions.
  • Novel Heading Estimation Techniques: To combat the degradation of PF performance due to angular noise (amplified by wheel slip), two methods are introduced:
  • Reinforcement Learning (RL) using Deep Q-Network (DQN): A trained neural network estimates the WMR's heading by taking the noisy location from RTLS and the PF's current weighted mean as input. This approach learns to incorporate real-world angular noise.
  • Location-based Heading Estimation (LHE): This method estimates the WMR's heading directly from its noisy location and the PF's weighted mean, without relying on wheel turning or motion direction. It uses the atan2 function to calculate the angle based on the relative positions.
  • Slip Estimation and Compensation: Once accurate localization and heading are achieved, the system uses the desired trajectory information and the PF's estimated position to estimate lateral slip along the X and Y axes. This slip information is then used to determine the direction the WMR needs to move to get back on track.

Datasets

The paper does not explicitly mention using pre-existing datasets for training or evaluation. Instead, it relies on simulations and experimental setups to generate data. The simulations are conducted in a virtual environment (Ubuntu 16.04, Python3, Keras with Tensorflow) and use Turtle graphics for visualization. The experiments involve a Wheeled Mobile Robot (WMR) equipped with a DecaWave system and anchors placed in a controlled workspace (e.g., 500 cm²). Noise is intentionally added to simulate real-world conditions, including distance noise (0-30 cm) and angular noise (0.1° increments).

Key Insights and Ideas

  • Angular noise is the primary culprit for PF degradation: The paper strongly emphasizes that while distance noise has a minimal impact, angular noise has an exponential and severe effect on PF localization accuracy, especially in environments with wheel slip.
  • Heading estimation is key to mitigating slip-induced localization errors: By accurately estimating the robot's heading, even in the presence of angular noise, the PF's localization performance can be dramatically improved.
  • RL and LHE offer robust heading estimation: Both proposed methods, RL and LHE, demonstrate significant improvements in localization accuracy by effectively handling angular noise. LHE, in particular, is highlighted for its simplicity and robustness.
  • Slip estimation is a direct consequence of accurate localization: Once the robot's position is reliably estimated, estimating and compensating for lateral slip becomes a tractable problem.
  • DecaWave's UWB technology is well-suited for indoor environments: The paper highlights the advantages of UWB for resource-constrained robots in indoor settings due to its low power consumption, robustness to multipath, and long-ranging capabilities.

Key Results

  • Distance noise impact: Average error in Particle Filter Estimated Position (PFEP) was less than 6 cm for distance uncertainties up to 25 cm.
  • Angular noise impact: Without angular noise, average error was around 9.93 cm. With cumulative angular noise (up to 54.6°), the error increased to 98.41 cm.
  • RL and LHE mitigation: Both RL and LHE methods significantly reduced the angular noise-induced localization error. For RL, the error was reduced to 9.78 cm, and for LHE, it was 14.99 cm.
  • Simulation validation: Simulation results consistently showed that the proposed RL and LHE heading estimation methods significantly improve PF localization and tracking performance on slippery surfaces.
  • Slip estimation accuracy: The paper demonstrates that lateral slip can be accurately estimated from PF location estimates and desired path information, enabling corrective actions.

Limitations and Future Work

  • Simulation-based validation: While simulations show promising results, the paper primarily relies on simulations for validation. Further real-world experiments with diverse terrains and more complex scenarios would strengthen the findings.
  • Workspace scaling: Although the RL model was trained on a 500 cm² space, the paper suggests it can be applied to larger spaces by scaling XY-coordinates. However, the practical implications and performance in significantly larger or irregularly shaped environments are not extensively explored.
  • Complexity of RL training: While RL is effective, training deep Q-networks can be computationally intensive and require careful hyperparameter tuning.
  • Future Work: The authors suggest that the proposed method can be applied to robots with any number of wheels and any shape, and in both indoor and outdoor environments. They also mention the potential to incorporate sensor measurement uncertainties into the slip estimation. The paper implies that this method can handle both linear and non-linear slip problems.

Summary updated

BibTeX
@article{ullah2018robust,
  author = {Zakir Ullah and Zhiwei Xu and Lei Zhang and Libo Zhang},
  title = {{A Robust Localization, Slip Estimation, and Compensation System for WMR in the Indoor Environments}},
  journal = {Symmetry},
  year = {2018},
  doi = {10.3390/sym10050149},
}
Figure