Estimation Procedures for Robust Sensor Control

Gregory D. Hager, Max MintzUAI 1987

TL;DR

This paper tackles the "sensor control problem" in robotics, where a robot must actively choose how to use its sensors to best estimate its environment, especially in the face of noisy and incomplete measurements. The authors propose and evaluate three estimation techniques—the Extended Kalman Filter (EKF), a discrete Bayes approximation, and an iterative Bayes approximation—demonstrating that the EKF can be unreliable in nonlinear systems, while Bayes approximations offer more robust performance.

What Problem Does This Paper Address?

Intelligent robots rely on accurate world models derived from sensor data. However, real-world sensors are imperfect: they have limited scope, introduce noise, and provide discrete information. This leads to uncertainty in the robot's understanding of its environment. Traditionally, robots passively accept sensor data, forcing conservative decision-making to account for potential errors. This paper argues for "active sensing," where robots strategically control their sensors to actively reduce uncertainty and improve estimation accuracy. The core challenge is the "sensor control problem": how to design an estimation procedure and a control strategy that minimizes estimation error while considering computational and time constraints. This is analogous to experimental design, where the goal is to maximize information gain within resource limitations.

Methods

The paper explores three estimation techniques for nonlinear measurement systems:

  • Extended Kalman Filter (EKF): A standard technique that linearizes a nonlinear system around a current estimate. The authors show that the EKF's performance can degrade significantly when the uncertainty in the parameter being estimated is large, leading to potential divergence.
  • Discrete Bayes Approximation: This method uses a finite set of "atoms" (possible values) for the parameter being estimated and computes the posterior distribution. The authors highlight that while this offers a more robust alternative to the EKF, its complexity increases significantly with the number of atoms and dimensions.
  • Iterative Bayes Approximation: This approach refines the discrete Bayes approximation by iteratively narrowing down the interval of uncertainty. When the estimation procedure starts to "quantize" (i.e., become less sensitive to new data), the algorithm focuses on a smaller interval, effectively improving accuracy at the cost of potentially more computation.

The paper also introduces a game-theoretic framework to analyze the robustness of estimation procedures against worst-case scenarios arising from linearization errors in the EKF.

Datasets

Not discussed in the paper. The evaluation is primarily based on simulations of a 2D stereo camera problem and mathematical analysis.

Key Insights and Ideas

  • EKF Fragility: The EKF, while widely used, is shown to be sensitive to the size of the uncertainty interval in nonlinear systems. Large uncertainties can lead to significant underestimation of actual errors and even filter divergence.
  • Bayes Approximations for Robustness: Discrete and iterative Bayes approximations offer a more robust alternative to the EKF, particularly in scenarios with significant uncertainty. They provide more predictable performance.
  • Game Theory for Robustness Analysis: Applying game theory provides a principled way to analyze the worst-case performance of estimation procedures, especially when dealing with uncertainties in linearization. This helps identify when an estimator might fail.
  • Iterative Refinement: The iterative Bayes approximation demonstrates the value of adaptive sampling and refinement. By focusing computation on smaller uncertainty regions, it can achieve better accuracy than a fixed-size batch approach.

Key Results

  • Simulations show that the Bayes approximation yields substantial performance improvements over the EKF, especially as the uncertainty interval widens. For instance, in one simulation, the Bayes approximation's observed mean square error was significantly lower than the EKF's.
  • The paper provides mathematical analysis demonstrating that the EKF's error terms can drastically under-represent the true estimation error when the uncertainty interval is large, leading to a failure to converge in practical scenarios.
  • Table 1 and Table 2 present quantitative comparisons, showing the EKF's error terms deviating significantly from observed errors as uncertainty increases, while the Bayes approximation shows more consistent performance.

Limitations and Future Work

  • Computational Complexity: The Bayes approximation, while robust, can be computationally expensive, especially for high-dimensional problems or when a very fine discretization is required.
  • Time/Accuracy Tradeoff: The paper acknowledges that a key question remains about the exact time/accuracy tradeoff between the EKF and finite Bayes techniques for a given number of samples.
  • Generalization: The authors plan to extend their game-theoretic analysis to multivariate cases and explore control strategies for sequential and batch estimation scenarios.
  • Multi-Sensor Systems: Future work includes generalizing these procedures for multi-sensor estimation and control, which will involve challenges in coordinating multiple sensors and arbitrating communication.
  • Structural Uncertainty: The paper also mentions considering cases where prior knowledge is structural rather than parametric, requiring reasoning about structural and environmental constraints.

Summary updated

BibTeX
@inproceedings{hager1987estimation,
  author = {Gregory D. Hager and Max Mintz},
  title = {{Estimation Procedures for Robust Sensor Control}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1987},
}
Figure