How can mobile sensors track targets with estimation-aware continuous-time optimal control?

Planning a trajectory that also shapes estimation uncertainty

Target tracking estimates a target system's state from measurements that are often collected by mobile sensors, such as aerial vehicles carrying range or bearing sensors. Because those measurements depend jointly on the sensor's state and the target's state, planning the sensor's path has to account for more than the sensor's own dynamics and operational limits: it must also account for how measurements collected along the trajectory affect estimation of the target.

The paper formulates an estimation-aware optimal control problem that minimizes a running cost over the mobile sensor state, the target covariance, and the control input, subject to the sensor's nonlinear dynamics, the covariance dynamics, fixed initial and final sensor states, state inequality constraints, and an input set. The planner is initialized with the latest target state estimate and covariance estimate supplied by the state estimator, so the trajectory is chosen with estimation quality in view rather than after the fact.

Why the covariance, not just observability, drives the plan

Existing target-tracking work uses several metrics to quantify how a sensor trajectory affects target-state estimation, including the observability Gramian, variants focused on the target's final state, and the Fisher information matrix. This paper uses the state-estimation error covariance, which directly quantifies the uncertainty of the state estimate and is attractive because widely used estimators such as the Kalman filter provide a covariance that accounts for process noise and measurement noise.

The method embeds the continuous-time extended Kalman filter. The filter propagates the mean and covariance of the target estimate, with the Kalman gain depending on the mobile sensor state, and the covariance differential equation coupling the target covariance to the sensor's position through the measurement Jacobian. Under the zero-innovation assumption used for planning, future measurements are unavailable, so the predicted measurement equals its expected value and the mean dynamics simplify to an integral of the target drift from the initial mean.

Constraints held between discretization points

A stated limitation of existing target-tracking methods is that they constrain the mobile sensor state only in discrete time. Constraints such as keep-out zones are often safety critical and must hold at all times, but informative sensing often requires aggressive maneuvers that drive the state close to constraint boundaries, so violations can occur between discretization points. Reducing those violations generally requires dense time discretization, which increases the size and computational cost of the trajectory optimization.

To address this, the paper transforms constraints on the mobile sensor into auxiliary continuous-time nonlinear dynamics with boundary conditions and augments the optimal control problem with an accumulated constraint violation term. The augmented state stacks the sensor state, the half-vectorized covariance, the accumulated violation, and the integral of the running cost, and the resulting problem is discretized and linearized for existing optimization algorithms. The paper reports that the method satisfies mobile-sensor constraints in continuous time within numerical precision even under sparse time discretization.

Exploiting covariance symmetry for efficient gradients

Enforcing continuous-time constraints introduces a computational challenge: gradient-based continuous-time optimal control differentiates the underlying dynamics equation, and in target tracking that equation introduces nonlinear coupling between the mobile sensor's state and the target covariance while also needing to preserve the covariance matrix's symmetric structure.

The paper provides an efficient formula that exploits the symmetry of the covariance matrix to differentiate the continuous-time nonlinear covariance dynamics in extended Kalman filtering. A half-vectorization transformation extracts the lower triangular entries of the covariance, including the diagonal, reducing the number of variables to n_p = n_ξ(n_ξ + 1)/2 and yielding a covariance-rate function used in the optimal control problem. The authors note that efficient differentiation formulas for the continuous-time extended Kalman filter covariance, and their use for continuous-time constraints in target tracking, had received little if any attention.

Two examples test the method under different dynamics

The method is demonstrated with the augmented Lagrangian algorithm on two examples. The first is LTE transmitter localization, with a fixed final time t_f = 10 and N = 5, initial sensor state x_0 = (1, 1, 0, 0), initial target state ξ_0 = (0, 0, 0.001), initial covariance P_0 = diag(4, 4, 1×10^-4), and final sensor state x_f = (0, 1, 0, 0). Constraint parameters include position bounds r_min_S = (-0.25, 0.75), r_max_S = (1.25, 1.25), velocity bounds v_min_S = -0.5 and v_max_S = 0.5, and input bounds u_min = (-0.2, -0.5) and u_max = (0.2, 0.5); noise covariances are Q = diag(1×10^-6, 1×10^-6, 1.4193×10^-4) and R = 5.

The second example is ground vehicle tracking, in which an unmanned aerial vehicle collects ranging measurements to a ground vehicle whose position, velocity, and heading are to be estimated. The target uses a unicycle model to showcase nonlinear target dynamics, with target state comprising position, forward velocity, angle from the x-axis, and angular velocity, and a scalar range measurement equal to the distance between sensor and target. Parameters are t_f = 10 and N = 5, initial sensor state x_0 = (0, 0, 0, 0), initial target state ξ_0 = (5, 5, 1, 0, 0.3), initial covariance P_0 = diag(10, 10, 0.5, 2, 0.1), final sensor state x_f = x_0, constraint parameters r_max_S = 4, r_min_ST = 1, and u_max = 2, noise covariances Q = diag(0.01, 0.01, 0.1, 0.1, 0.1) and R = 5.

Weighting covariance against energy moves the estimate

The results compare three weighting schemes: covariance only (ω_p = 1), energy only (ω_u = 1), and balanced (ω_p = ω_u = 0.5). Figure 1 tracks the Schatten 2-norm of the error-covariance matrix, ||P(t)||_2, over time for both examples. As the covariance weight ω_p increases, the area under the covariance trajectory decreases, indicating improved estimation performance relative to the minimum-energy solution at ω_p = 0.

Figure 2 shows Pareto fronts for the trade-off between the energy and covariance objectives when sweeping the objective weights ω_p and ω_u, for both examples. These fronts make the trade-off explicit: the two objectives are not aligned, and the choice of weights determines where the planned trajectory lands between estimation quality and control effort. The paper's own framing is that the method achieves estimation-aware target tracking while satisfying mobile-sensor constraints in continuous time, even under sparse time discretization.

What the evidence covers and what it does not

The evidence for continuous-time constraint satisfaction and estimation-aware tracking comes from numerical simulations of two examples, LTE transmitter localization and ground vehicle tracking, both solved with the augmented Lagrangian method at N = 5. The paper's comparison is against off-the-shelf solvers that only enforce constraints at discretization points; the claim of continuous-time satisfaction is stated within numerical precision, and the planning formulation relies on the zero-innovation assumption because future measurements are unavailable during path planning.

The method is presented for nonlinear sensor dynamics, nonlinear target dynamics, nonlinear measurement models, and nonlinear covariance dynamics in extended Kalman filtering, so its scope is the class of problems captured by those models and the stated constraint and noise parameters. The paper does not report hardware experiments, and the reported outcomes are simulation results under the listed initial conditions, constraint parameters, and noise covariances.

The priority is constraint timing, not just covariance weight

The paper's distinctive priority is the timing at which constraints are enforced. Covariance weighting is a tunable trade-off, and the Pareto fronts show that increasing ω_p lowers the area under the covariance trajectory relative to the minimum-energy solution, but the method's differentiating claim is that sensor state and input constraints hold in continuous time within numerical precision while still supporting sparse time discretization.

For a reader deciding whether this matters, the relevant comparison is between a planner that enforces constraints only at discretization points and one that augments the problem with an accumulated violation term and auxiliary continuous-time constraint dynamics. The paper's evidence supports the latter under the two simulated scenarios and the stated parameters; the choice of ω_p and ω_u then sets where on the covariance-energy trade-off the resulting trajectory sits.

Sources used in this answer