Model-Based Learning of Turbulent Flows using Mobile Robots

Reza Khodayi-mehr, Michael M. ZavlanosarXiv 2018 · OpenReview

TL;DR

This paper introduces a novel physics-based learning framework for mobile robots to accurately map complex turbulent flow fields. By combining offline physics simulations with real-time data collection and active learning, the system significantly reduces the number of measurements needed compared to purely data-driven approaches, demonstrating improved accuracy and robustness in both simulation and real-world experiments.

What Problem Does This Paper Address?

Accurately sensing environmental fields (EFs), such as turbulent flows, is crucial for many robotic applications, from underwater exploration to atmospheric monitoring. Traditional purely data-driven methods, like those using Gaussian Processes (GPs), often require an impractically large number of measurements to learn complex EFs. On the other hand, physics-based models (e.g., RANS equations) offer global knowledge but suffer from uncertainties in parameters, require experimental validation, and are computationally too demanding for onboard mobile robots. This paper bridges this gap by proposing a method that leverages the strengths of both approaches, enabling accurate and efficient environmental sensing with mobile robots.

Methods

The core of the proposed method is a Bayesian framework for model selection.

  1. Physics-Based Prior Generation: Offline, a pool of potential physics-based models (using RANS equations) for the turbulent flow field is generated by varying uncertain parameters (like boundary conditions). These models, though potentially inconsistent with each other or reality, inform the prior mean of Gaussian Process (GP) models for the flow properties (mean velocity and turbulent intensity).
  2. Gaussian Processes (GPs): GPs are used to construct statistical models for the flow properties. The prior mean is informed by the RANS simulations, and the prior uncertainty is derived from RANS-estimated turbulent intensity, providing a more informed starting point than uninformative priors.
  3. Bayesian Model Selection: As the mobile robot collects measurements, a Bayesian approach updates the probabilities of the pre-generated physics-based models. This allows the system to select the most likely model(s) that best explain the observed data.
  4. Active Learning for Data Acquisition: A custom-built mobile robot is controlled using an active learning strategy. This strategy guides the robot to waypoints that maximize the information content of the measurements, specifically by minimizing the posterior entropy of unobserved areas. This ensures efficient data collection.
  5. Uncertainty Quantification: The framework explicitly models and accounts for various sources of uncertainty, including sensor noise, robot heading errors, and measurement location errors, deriving accurate expressions for measurement noise.

Datasets

  • Simulated Data: A simulated 10x10 m² office environment was used to generate ground truth data for evaluating the framework's performance.
  • Experimental Data: Real-world experimental data was collected using a custom-built mobile robot equipped with OMRON D6F-W01A1 flow sensors in a 2.2 × 2.2 × 0.4m³ test rig. This data was used to demonstrate the framework's robustness to real-world uncertainties.

Key Insights and Ideas

  • Physics-Informed Priors are Crucial: Using physics-based models (even with uncertainties) to inform the prior of statistical models (like GPs) dramatically improves learning efficiency and accuracy compared to purely data-driven methods with uninformative priors.
  • Active Learning Maximizes Information Gain: Strategically choosing measurement locations based on an entropy-driven metric significantly outperforms random or lattice-based sampling, leading to faster convergence and more accurate field estimation.
  • Bayesian Model Selection Handles Uncertainty: The framework's ability to assign probabilities to different physics-based models allows it to adapt to real-world uncertainties and select the most plausible model for the observed flow.
  • Turbulent Intensity as a Prior Uncertainty Source: Incorporating RANS-estimated turbulent intensity into the GP kernel provides a more realistic prior uncertainty, improving the model's ability to capture local flow variations.
  • Experimental Validation is Key: Demonstrating the framework's effectiveness on a real robot in a controlled environment, accounting for practical noise sources, is a significant step beyond purely simulation-based studies.

Key Results

  • Improved Prediction Accuracy: In simulations, the proposed physics-based method achieved a three-fold improvement in prediction error compared to a purely data-driven approach. The posterior prediction error was 328% lower than that of a purely data-driven approach.
  • Efficient Data Collection: The active learning strategy significantly reduced the number of measurements required to achieve accurate flow field estimations.
  • Robustness to Real-World Uncertainties: Experimental results showed that the framework correctly selects models that agree with empirical data and reaches a steady state with reasonable uncertainty bounds.
  • Model Selection Success: In experiments, the framework successfully identified the most likely physical model (RSM model 7) from a pool of 12 models, with its posterior probability converging to 1.
  • Uncertainty Quantification: The framework provided reasonable one-standard-deviation uncertainty bounds for its predictions, with 90% of measurements falling within these bounds.

Limitations and Future Work

  • Assumption of Stationarity: The current framework assumes statistically steady and ergodic turbulent flows. Extending it to handle time-dependent (unsteady) turbulent flows is a future research direction, though challenging due to the difficulty in empirically measuring and modeling time-varying properties.
  • Model Availability: The framework relies on the availability of reliable physical models and domain geometry. Its performance is limited if the prior models are uninformative.
  • Correlation of Velocity Components: The current work assumes independence between velocity components for simplicity. Future work could incorporate correlations using multi-output GPs.
  • Turbulent Intensity Modeling: While GPs were used for turbulent intensity, more advanced probabilistic models could potentially improve performance.
  • Scalability: While the framework handles large domains efficiently, further investigation into distributed implementations for teams of mobile robots is planned.
  • Correlation Length Scale: Using a constant correlation length scale in the GP kernel is a simplification. Future work could explore making this scale a function of flow characteristics or learning it from data.

Summary updated

Abstract

We propose a physics-based method to learn environmental fields (EFs) using a mobile robot. Common purely data-driven methods require prohibitively many measurements to accurately learn such complex EFs. Alternatively, physics-based models provide global knowledge of EFs but require experimental validation, depend on uncertain parameters, and are intractable for mobile robots. To address these challenges, we propose a Bayesian framework to select the most likely physics-based models of EFs in real-time, from a pool of numerical solutions generated offline as a function of the uncertain parameters. Specifically, we focus on turbulent flow fields and utilize Gaussian processes (GPs) to construct statistical models for them, using the pool of numerical solutions to inform their prior mean. To incorporate flow measurements into these GPs, we control a custom-built mobile robot through a sequence of waypoints that maximize the information content of the measurements. We experimentally demonstrate that our proposed framework constructs a posterior distribution of the flow field that better approximates the real flow compared to the prior numerical solutions and purely data-driven methods.
BibTeX
@misc{khodayimehr2018model,
  author = {Reza Khodayi-mehr and Michael M. Zavlanos},
  title = {{Model-Based Learning of Turbulent Flows using Mobile Robots}},
  year = {2018},
  eprint = {1812.03894},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1812.03894},
}
Figure