Discovering and exploiting active sensing motifs for estimation

This paper introduces the BOUNDS method and the Augmented Information Kalman Filter (AI-KF) to systematically discover active sensing motifs and dynamically fuse neural network with model-based estimation, thereby improving state estimation for nonlinear autonomous systems in GPS-denied environments.

The Gist

Imagine you are trying to navigate a foggy forest with a broken compass. You can't see the trees clearly, and you don't know exactly where you are. If you just stand still or walk in a straight line, you might never figure out your location or the direction of the wind. But, if you start dancing—spinning in circles, stopping suddenly, or running back and forth—you might catch a glimpse of a landmark or feel a change in the wind that tells you exactly where you are.

This is the core idea of Active Sensing: moving your sensors (or your body) on purpose to gather better information.

This paper introduces two new tools to help robots and scientists master this "dance": BOUNDS and the AI-KF.

1. The Problem: The "Blind" Robot

Autonomous systems (like drones or self-driving cars) and even animals (like flies or mice) need to guess things they can't see directly, like their speed, altitude, or wind direction.

• The Challenge: In the real world, things are messy. Sensors are noisy, and sometimes the math says, "You can't figure this out right now."
• The Old Way: Engineers usually try to build better sensors or use standard math filters (like the Kalman Filter). But these often fail if the robot starts in the wrong place or if the data is too confusing. They are like a student who memorizes a formula but panics when the test question is slightly different.

2. Tool #1: BOUNDS (The "Detective Map")

BOUNDS stands for Bounding Observability for Uncertain Nonlinear Dynamic Systems. That's a mouthful, so let's call it the "Detective Map."

• What it does: It simulates a robot moving in a computer and asks, "If I do this specific move (like a sharp turn), how much clearer does my picture of the world become?"
• The Analogy: Imagine you are trying to guess the shape of a giant, invisible statue in a dark room by feeling it with your hands.
  • If you just stand still and touch one spot, you might think it's a flat wall.
  • BOUNDS is like a smart guide that tells you: "Don't just stand there! Walk to the left and touch the top. Now you'll realize it's a sphere!"
• The Discovery: The researchers found that different "moves" reveal different secrets.
  • To figure out wind direction, the robot needs to turn its head (change its heading).
  • To figure out altitude (how high it is), the robot just needs to speed up or slow down in a straight line.
  • There is no "one size fits all" move. You have to pick the right dance step for the specific question you are trying to answer.

3. Tool #2: The AI-KF (The "Smart Translator")

Once the robot knows when it has good information (thanks to BOUNDS), it needs a way to use it. Enter the Augmented Information Kalman Filter (AI-KF).

• The Problem with Old Filters: Traditional filters are like a strict teacher who only listens to you if you already know the answer. If you guess wrong at the start, the teacher gets confused and never corrects you. Also, they only look at the current moment, forgetting the history of how you got there.
• The AI-KF Solution: This is a hybrid system that combines two brains:
  1. The Model Brain (The Traditional Filter): Uses physics and math to predict where the robot should be.
  2. The Data Brain (Neural Network): A "black box" AI that looks at a long history of sensor data to guess where the robot is, even if it started with a bad guess.
• The Magic Trick: The AI-KF uses the "Detective Map" (BOUNDS) to decide who to trust.
  • Scenario: The robot is flying straight (bad visibility). The AI-KF says, "The Data Brain is guessing wildly right now. Ignore it and stick to the Model Brain."
  • Scenario: The robot suddenly accelerates (good visibility!). The AI-KF says, "The Data Brain just saw something amazing! Trust it immediately and update our position!"
• The Result: The robot can recover from bad starts and handle messy, real-world data much better than before. It's like having a navigator who knows exactly when to ignore the GPS signal and when to trust it completely.

4. Real-World Proof: The Drone Test

The team tested this on a real drone (quadcopter) flying outside without GPS.

• The Setup: The drone had to guess its height and speed using only a camera looking down and an accelerometer (a vibration sensor).
• The Result:
  • Old Method (Standard Filter): If the drone started with the wrong guess for its height, it got stuck in a loop of errors and crashed or flew wildly.
  • New Method (AI-KF): Even if the drone started with a terrible guess, as soon as it did a little acceleration maneuver, the system realized, "Aha! Now I can see!" and instantly corrected its course to the right height.

Why This Matters

This research bridges the gap between biology and engineering.

• For Nature: It explains why animals (like flies) wiggle and turn so much. They aren't just being chaotic; they are performing specific "active sensing" moves to unlock information about the wind and their position.
• For Robots: It allows us to build robots that need fewer sensors (saving money and weight) but are smarter because they know how to move to get the information they need.

In short: Instead of building a robot with a million perfect cameras, we can build a robot with a few cheap sensors that knows how to "dance" to see the world clearly.

Technical Summary

Here is a detailed technical summary of the paper "Discovering and exploiting active sensing motifs for estimation" by Cellini et al.

1. Problem Statement

Autonomous systems (both biological and engineered) rely on sensory cues to estimate unknown states about themselves or their environments. A fundamental challenge arises in nonlinear systems where variables of interest are nonlinearly related to sensor measurements (e.g., optic flow encodes velocity-to-distance ratios, making independent estimation impossible from a single static moment).

• Active Sensing: The strategy of moving sensors or the agent to decouple correlated variables and enhance estimation is widespread in biology but underutilized in engineering.
• The Gap: Existing methods lack the ability to systematically discover which specific movements (motifs) improve the estimation of specific state variables in partially observable, nonlinear systems using experimental data.
• Estimation Instability: Classical estimators (like the Kalman Filter) often fail when observability is time-varying or weak, particularly if initial state guesses are poor or if information is only available sporadically during specific maneuvers.

2. Methodology

The authors propose a two-part framework: BOUNDS for discovering active sensing strategies and the AI-KF for exploiting them during state estimation.

A. BOUNDS (Bounding Observability for Uncertain Nonlinear Dynamic Systems)

BOUNDS is an empirical computational pipeline designed to quantify the observability of individual state variables along dynamic trajectories.

• Inputs: A time-series state trajectory (simulated or real) and a system model (dynamics $f$ and measurements $h$).
• Process:
  1. Trajectory Reconstruction: Uses Model Predictive Control (MPC) to determine control inputs that reconstruct the observed trajectory if inputs are unknown.
  2. Empirical Jacobian Calculation: Perturbs each initial state variable $x_{i,0}$ by a small amount $\epsilon$ in positive and negative directions while holding control inputs constant. It records the resulting change in measurements over a time window $\omega$.
  3. Observability Matrix ($O$): Constructs the empirical Jacobian $\Delta Y / \Delta x_0$, representing the sensitivity of measurements to state perturbations.
  4. Fisher Information & Cramér-Rao Bound: Computes the Fisher Information Matrix $F = O^T R^{-1} O$ (where $R$ is sensor noise covariance).
  5. Regularized Inverse: Calculates the inverse $F^{-1}$ using a "Chernoff inverse" approach ($F + \lambda I$) to handle singular matrices in partially observable systems. The diagonal elements of $F^{-1}$ represent the minimum error variance for each state variable.
  6. Motif Discovery: By sliding this window along a trajectory, BOUNDS reveals which movement patterns (e.g., turns, accelerations) correlate with spikes in observability (drops in error variance) for specific states.

B. Augmented Information Kalman Filter (AI-KF)

The AI-KF is a hybrid estimator that fuses model-based filtering with data-driven estimates, weighted by observability.

• Data-Driven Estimator ($H_i$): A feed-forward Artificial Neural Network (ANN) trained to estimate a state variable (e.g., wind direction) using a history of measurements. Crucially, the ANN is trained on data curated by observability levels (focusing on high-observability maneuvers) to ensure accuracy during active sensing bouts.
• Observability Estimator ($G_i$): An ANN or heuristic model that predicts the current observability (or error variance) of the state based on current measurements and inputs.
• Fusion Mechanism:
  • The standard Kalman Filter measurement vector is augmented with the ANN's state estimate.
  • The measurement noise covariance matrix ($R$) is dynamically updated. The variance assigned to the ANN's estimate is inversely proportional to the predicted observability (derived from $G_i$).
  • Logic: When observability is high (e.g., during a turn), the AI-KF trusts the ANN estimate heavily. When observability is low (e.g., straight flight), the filter relies on the model-based prediction, preventing the filter from diverging due to poor ANN performance.
  • Double-Counting Prevention: A "relevance ratio" mechanism ensures that if the ANN estimate and the KF prediction are statistically indistinguishable, the augmented information is suppressed to avoid artificially low error variance.

3. Key Contributions

1. BOUNDS Framework: A novel, model-agnostic method to quantify observability for individual state variables in nonlinear, partially observable systems using experimental or simulated data. It moves beyond binary "observable/unobservable" classifications to provide quantitative error bounds.
2. Discovery of State-Specific Motifs: Demonstrated that different state variables require distinct active sensing strategies. For example, wind direction estimation requires heading changes, while altitude estimation can be achieved via straight-line acceleration.
3. AI-KF Architecture: A theoretically grounded hybrid estimator that dynamically weights data-driven and model-based estimates based on real-time observability. It overcomes the sensitivity of Kalman Filters to poor initialization and sparse observability.
4. Python Package (pybounds): An open-source tool to facilitate the application of these methods to other systems.

4. Results

The framework was validated on a 3D quadcopter model and experimental data from a GPS-denied flight.

• Active Sensing Motifs (Simulation):

  • Wind Direction: Observable only during heading changes (turns). Straight flight provided no information.
  • Altitude: Observable during acceleration/deceleration and offset turns, provided optic flow magnitude is available.
  • Ground Speed: Estimable during turns or acceleration depending on sensor availability (e.g., adding apparent airflow magnitude enabled estimation during offset turns).
  • Sensor Dependency: The paper quantified exactly which sensor combinations (e.g., heading + airflow vs. optic flow + acceleration) were sufficient for specific states.

• Estimation Performance:

  • ANN Training: Training ANNs on data sorted by high observability significantly reduced estimation error compared to random training data.
  • AI-KF vs. Standard UKF:
    • In simulations with poor initial state guesses and biased noise, the standard Unscented Kalman Filter (UKF) often failed to converge or diverged.
    • The AI-KF converged rapidly to the true state during high-observability maneuvers (accelerations/turns) and maintained stability during low-observability periods.
    • The AI-KF showed robustness to variations in process noise ($Q$) and initial covariance ($P_0$), whereas the UKF performance was highly sensitive to these hyperparameters.
  • Real-World Validation: Applied to a DJI Matrice 300 RTK quadcopter flying outdoors without GPS. The AI-KF successfully estimated altitude and forward velocity using only ventral optic flow and accelerometers, converging within seconds where the standard UKF failed to converge over a 50-second trajectory.

5. Significance

• For Engineering: Provides a systematic design tool for creating sensor-minimal autonomous systems. Engineers can use BOUNDS to determine the minimum sensor suite and specific maneuvering strategies required to estimate critical states, reducing hardware costs and complexity.
• For Biology: Offers a quantitative framework to test hypotheses about animal behavior. It suggests that animals may perform specific maneuvers (like turns or accelerations) not just for navigation, but to "reset" or improve the observability of specific environmental variables (e.g., wind direction).
• For Estimation Theory: Bridges the gap between classical model-based control and modern data-driven methods. The AI-KF provides a statistically rigorous way to integrate neural networks into control loops without sacrificing stability guarantees, specifically addressing the "sporadic information" problem inherent in active sensing.

In summary, this work transforms active sensing from a heuristic concept into a quantifiable, designable engineering principle, enabling more robust and efficient autonomous systems in complex, nonlinear environments.