betaFeedback

WorksComputer Science

Disturbance-Aware Informative Planning in Spatiotemporal Environments

Abstract

Autonomous robots are widely used for persistent monitoring of spatiotemporal environments, where quantities of interest (QoI) change across space and time and require repeated measurements. In practice, robot motion can be affected by external disturbances that vary across the environment. When such external disturbances are not considered in trajectory optimization, it can lead to inaccurate predictions of robot motion and, consequently, the information acquired about the QoI during execution. However, estimating disturbances throughout the environment can be inefficient, as disturbances in regions with little overlap with monitoring trajectories have limited impact on monitoring performance. We therefore propose a two-stage informative planning framework for unknown, spatially varying but time-invariant disturbances, combining informed disturbance sampling with disturbance-aware trajectory optimization. First, we identify where QoI measurements are needed and generate alternative paths to these locations using the environment geometry. We then prioritize disturbance measurements based on their expected reduction of disturbance uncertainty along these paths. Second, rather than optimizing information acquisition based on nominal robot motion, we incorporate the learned disturbance belief into trajectory optimization to account for its effects on robot motion and information acquisition. Across four simulation scenarios, the proposed method reduces disturbance-learning time by 25.0–52.3 percent compared with environment-wide estimation and identifies paths that appear costly under nominal dynamics but become preferable when disturbances are considered.

Introduction

Autonomous robots are widely used for environmental monitoring, collecting measurements of quantities of interest such as temperature, water quality, and terrain characteristics [6, 3, 18]. In spatiotemporal environments, these quantities of interest evolve over time, causing information decay. Informative planning must therefore account for both information acquisition and decay when optimizing robot trajectories. A separate challenge is uncertainty about external disturbances, which can affect robot motion and the locations where measurements are collected. Consequently, trajectories optimized using nominal dynamics may acquire less information than predicted during execution.

Addressing these challenges requires accounting for uncertainty about the quantities of interest being monitored and uncertainty about the external disturbances affecting robot motion. Informative planning methods quantify uncertainty in the quantities of interest using measures such as posterior variance, entropy, mutual information, and clarity [9, 19, 1]. These measures guide the selection of sensing locations and trajectories through sampling-based planning under travel budgets [19, 7] or optimization of expected information gain [6, 18, 12]. In spatiotemporal environments, informative planning must account for evolving quantities of interest. Existing approaches use spatiotemporal Gaussian processes [3], update sparse spatial GP models online [11], or explicitly model information acquisition and decay in ergodic exploration and trajectory optimization [14, 15]. However, they do not jointly address disturbance learning and its effects on robot motion during informative planning.

Beyond uncertainty in the quantities of interest, other approaches account for the effects of disturbances on robot motion. Popovic et al. [17] consider localization and mapping uncertainty in informative planning, while Cho et al. [5] incorporate known wind through direction-dependent travel costs. For unknown disturbances, Sun et al. [20] use Gaussian processes to learn state-dependent disturbances for safe exploration. Other approaches combine online wind estimation with wind-aware routing [13] or use learned wind maps in quadrotor planning and control to reduce attitude effort [21]. While these methods address disturbance estimation or its effects on robot motion, they do not jointly consider monitoring-driven disturbance learning and the incorporation of learned disturbance beliefs into informative trajectory optimization.

Selecting disturbance measurements to improve subsequent monitoring requires considering how reducing disturbance uncertainty affects future planning decisions. This is closely related to dual control, which balances immediate performance and learning [2, 10], and task-optimal exploration, which prioritizes learning according to its effect on task performance [23, 22]. However, directly evaluating this benefit is computationally challenging, as it requires considering possible measurement outcomes, their effects on the disturbance belief, and the resulting changes in subsequent informative trajectories.

To address these challenges, we propose a two-stage disturbance-aware informative planning framework. (i) The first stage predicts where QoI measurements are needed as information decays and generates alternative paths to these locations using environment geometry. We then prioritize disturbance measurements by their expected reduction of disturbance uncertainty along these paths, weighted by monitoring relevance and accounting for travel cost. (ii) The second stage incorporates the learned disturbance belief into Model-Based Diffusion (MBD) trajectory optimization, so that the trajectory cost reflects information acquisition and decay along predicted robot motion under disturbances rather than nominal motion. We compare against nominal planning, disturbance-aware monitoring without dedicated learning, and environment-wide disturbance estimation, evaluating monitoring performance over both stages.

Preliminaries

Notation

Let N={0,1,…}\mathbb{N} = \{0,1,\ldots\} and N>0={1,2,…}\mathbb{N}_{>0} = \{1,2,\ldots\}; R\mathbb{R}, R≥0\mathbb{R}_{\geq0}, and R>0\mathbb{R}_{>0} denote the real, nonnegative, and positive real numbers. For k∈N>0k \in\mathbb{N}_{>0}, IkI_k, S+k\mathbb{S}^{k}_{+}, and S++k\mathbb{S}^{k}_{++} denote the identity matrix and the sets of symmetric positive-semidefinite and positive-definite matrices in Rk×k\mathbb{R}^{k \times k}, respectively. We write N(μ,Σ)\mathcal{N}(\mu, \Sigma) for a Gaussian distribution with mean μ\mu and covariance Σ\Sigma, tr⁡(⋅)\operatorname{tr}(\cdot) for the trace, [a]+=max⁡{a,0}[a]_{+} = \max\{a, 0\} for a∈Ra \in\mathbb{R}, and IA(z)\mathbb{I}_{A}(z) for the indicator of z∈Az \in A.

Dynamics Model

Consider the continuous-time robot dynamics

x˙=f(x,u,dˉ(Ψ(x))),t≥t0,(1)\dot{x} = f\left(x, u, \bar{d}(\Psi(x))\right), \qquad t \ge t_{0}, \tag*{(1)}

where x∈X⊂Rnx \in\mathcal{X} \subset\mathbb{R}^{n} and u∈U⊂Rmu \in\mathcal{U} \subset\mathbb{R}^{m} denote the robot state and control input. The robot’s position is given by Ψ:X→P\Psi: \mathcal{X} \to\mathcal{P}, where P⊂Rnr\mathcal{P} \subset\mathbb{R}^{n_r} is the workspace and nr∈N>0n_r \in\mathbb{N}_{>0}. The unknown disturbance field is dˉ:P→Rnd\bar{d} : \mathcal{P} \to\mathbb{R}^{n_d}, with nd∈N>0n_d \in\mathbb{N}_{>0}, and f:X×U×Rnd→Rnf : \mathcal{X} \times\mathcal{U} \times\mathbb{R}^{n_d} \to\mathbb{R}^{n} is the known dynamics model. The field dˉ\bar{d} can vary with position, but its value at any fixed r∈Pr \in\mathcal{P} does not change with time. Let S⊂X\mathcal{S} \subset\mathcal{X} denote the prescribed set of admissible states, including obstacle avoidance.

Definition 1 (Trajectory). Let T=[ti,tf]⊂R\mathcal{T} = [t_i, t_f] \subset\mathbb{R}, where ti<tft_i < t_f. A trajectory starting from (ti,xi)∈R×X(t_i, x_i) \in\mathbb{R} \times\mathcal{X} is a pair of functions ξ=(ξx:T→X,ξu:T→U)\xi= (\xi_x : \mathcal{T} \to\mathcal{X}, \xi_u : \mathcal{T} \to\mathcal{U}) satisfying ξ˙x(t)=f(ξx(t),ξu(t),dˉ(Ψ(ξx(t))))\dot{\xi}_x(t) = f\left(\xi_x(t), \xi_u(t), \bar{d}(\Psi(\xi_x(t)))\right), ∀t∈(ti,tf)\forall t \in(t_i, t_f), with ξx(ti)=xi\xi_x(t_i) = x_i.

Quantity of Interest Model and Clarity

Let Np∈N>0N_p \in\mathbb{N}_{>0} be the number of free-space monitoring cells and Pcells={1,…,Np}\mathcal{P}_{\mathrm{cells}} = \{1, \ldots, N_p\} their index set. Cell p∈Pcellsp \in\mathcal{P}_{\mathrm{cells}} has center rp∈Pr_p \in\mathcal{P} and a scalar quantity of interest mp(t)∈Rm_p(t) \in\mathbb{R}. The cell quantities have independent Gaussian initial beliefs. Their process and observation models are

dmp(t)=Qp dWp(t),dyp(t)=Cp(x)mp(t) dt+R dVp(t),(2)\begin{aligned} dm_p(t) &= \sqrt{Q_p}\,dW_p(t),\\ dy_p(t) &= C_p(x)m_p(t)\,dt + \sqrt{R}\,dV_p(t), \tag*{(2)} \end{aligned}

where Qp∈R≥0Q_p \in\mathbb{R}_{\ge0} and R∈R>0R \in\mathbb{R}_{>0} are process- and measurement-noise intensities. The scalar processes Wp,VpW_p, V_p are mutually independent standard Wiener processes, and yp(t)∈Ry_p(t) \in\mathbb{R} is the integrated observation. The function Cp:X→RC_p : \mathcal{X} \to\mathbb{R} specifies the measurement sensitivity to cell pp and is zero when that cell is not observed. Different values of QpQ_p allow the quantities at different cells to evolve with different process uncertainty.

Estimating each mpm_p with a continuous-time Kalman filter gives posterior mean m^p(t)∈R\hat{m}_p(t) \in\mathbb{R} and variance Pp(t)∈R≥0P_p(t) \in\mathbb{R}_{\ge0}, with P˙p=Qp−Cp(x)2Pp2/R\dot{P}_p = Q_p - C_p(x)^2P_p^2/R. Clarity [1] is a normalized representation of differential entropy; for this scalar Gaussian belief, qp=(1+Pp)−1∈(0,1]q_p = (1 + P_p)^{-1} \in(0, 1]. Substitution of the variance equation yields

q˙p=gp(x,qp):=Cp(x)2R(1−qp)2−Qpqp2,gp:X×[0,1]→R.(3)\dot{q}_p = g_p(x, q_p) := \frac{C_p(x)^2}{R}(1 - q_p)^2 - Q_pq_p^2, \qquad g_p : \mathcal{X} \times[0, 1] \to\mathbb{R}. \tag*{(3)}

The first term represents information acquisition and the second represents information decay. Let qˉp∈[0,1]\bar{q}_p \in[0, 1] be the target clarity in cell pp. With q(t)=[qp(t)]p=1Np∈(0,1]Npq(t) = [q_p(t)]_{p=1}^{N_p} \in(0, 1]^{N_p}, the cellwise deficits and mean-deficit function qd:(0,1]Np→[0,1]q_d : (0, 1]^{N_p} \to[0, 1] are

δp(t)=[qˉp−qp(t)]+∈[0,1],qd(q(t))=1Np∑p=1Npδp(t)∈[0,1].(4)\delta_p(t) = [\bar{q}_p - q_p(t)]_{+} \in[0, 1], \qquad q_d(q(t)) = \frac{1}{N_p}\sum_{p=1}^{N_p}\delta_p(t) \in[0, 1]. \tag*{(4)}

Thus, δp\delta_p measures the difference between target and current clarity.

Disturbance Belief

To estimate the unknown disturbance field dˉ\bar{d}, the robot collects noisy measurements along its trajectory. Let tjd≥t0t_j^d \ge t_0 denote the time of the jj-th disturbance measurement, where j∈N>0j \in\mathbb{N}_{>0} indexes successive measurements. At position rj=Ψ(x(tjd))∈Pr_j = \Psi(x(t_j^d)) \in\mathcal{P}, the measurement yjd∈Rndy_j^d \in\mathbb{R}^{n_d} is given by

yjd=dˉ(rj)+νjd,νjd∼N(0,Rd),(5)y_j^d = \bar{d}(r_j) + \nu_j^d,\qquad\nu_j^d \sim\mathcal{N}(0,R_d), \tag*{(5)}

where νjd∈Rnd\nu_j^d \in\mathbb{R}^{n_d} is the measurement noise with covariance Rd∈S++ndR_d \in\mathbb{S}_{++}^{n_d}. The measurement noises are independent across jj and independent of the noises in (2). The dataset Dtd={(rj,yjd):t0≤tjd≤t}\mathcal{D}_t^d = \{(r_j,y_j^d): t_0 \le t_j^d \le t\} contains all disturbance measurements collected up to time tt.

To represent uncertainty about the unknown field dˉ\bar{d}, we consider a random field d:P→Rndd:\mathcal{P}\to\mathbb{R}^{n_d}. Let Btd\mathcal{B}_t^d denote its posterior belief conditioned on Dtd\mathcal{D}_t^d, and let Bd\mathcal{B}_d be the space of such disturbance beliefs. For any r,r′∈Pr,r' \in\mathcal{P}, the posterior mean and cross-covariance are

μd(r,t)=EBtd[d(r)]∈Rnd,\mu_d(r,t) = \mathbb{E}_{\mathcal{B}_t^d}[d(r)] \in\mathbb{R}^{n_d},
Σd(r,r′,t)=Cov⁡Btd[d(r),d(r′)]∈Rnd×nd.(6)\Sigma_d(r,r',t) = \operatorname{Cov}_{\mathcal{B}_t^d}[d(r),d(r')] \in\mathbb{R}^{n_d \times n_d}. \tag*{(6)}

The marginal covariance is Pd(r,t)=Σd(r,r,t)∈S+ndP_d(r,t) = \Sigma_d(r,r,t) \in\mathbb{S}_{+}^{n_d}. Although the physical disturbance field dˉ\bar{d} is time-invariant, its posterior belief Btd\mathcal{B}_t^d is updated as new measurements are collected.

Probabilistic Online Attentive Mapping (POAM) for Estimating Disturbance

We estimate the unknown disturbance field using Probabilistic Online Attentive Mapping (POAM) [4], an online sparse Gaussian-process (GP) method. Its attentive kernel combines base kernels of different lengthscales using location-dependent weights, allowing spatially varying correlations. We approximate Btd\mathcal{B}_t^d by independent sparse GP posteriors for the disturbance components and assume diagonal observation-noise covariance RdR_d.

For component da:P→Rd_a:\mathcal{P}\to\mathbb{R}, a∈{1,…,nd}a\in\{1,\ldots,n_d\}, we use a zero-mean GP prior with kernel ka,t:P×P→Rk_{a,t}:\mathcal{P}\times\mathcal{P}\to\mathbb{R}, whose parameters are updated as new measurements become available. Let Ma∈N>0M_a\in\mathbb{N}_{>0} and Za,t=(za,1,t,…,za,Ma,t)∈PMaZ_{a,t}=(z_{a,1,t},\ldots,z_{a,M_a,t})\in\mathcal{P}^{M_a} denote the number and locations of inducing points. The corresponding inducing values ζa,t=[da(za,1,t),…,da(za,Ma,t)]⊤∈RMa\zeta_{a,t}=[d_a(z_{a,1,t}),\ldots,d_a(z_{a,M_a,t})]^\top\in\mathbb{R}^{M_a} have variational posterior N(ma,t,Sa,t)\mathcal{N}(m_{a,t},S_{a,t}), where ma,t∈RMam_{a,t}\in\mathbb{R}^{M_a} and Sa,t∈S++MaS_{a,t}\in\mathbb{S}_{++}^{M_a}. Define the nonsingular kernel matrix Ka,t=ka,t(Za,t,Za,t)∈S++MaK_{a,t}=k_{a,t}(Z_{a,t},Z_{a,t})\in\mathbb{S}_{++}^{M_a} and ha,t(r)=Ka,t−1ka,t(Za,t,r)∈RMah_{a,t}(r)=K_{a,t}^{-1}k_{a,t}(Z_{a,t},r)\in\mathbb{R}^{M_a}, where ka,t(Za,t,r)∈RMak_{a,t}(Z_{a,t},r)\in\mathbb{R}^{M_a} contains the kernel values between the inducing points and rr. For any r,r′∈Pr,r'\in\mathcal{P}, the predictive mean and covariance are

μd,a(r,t)=ha,t(r)⊤ma,t,\mu_{d,a}(r,t)=h_{a,t}(r)^\top m_{a,t},
Σd,a(r,r′,t)=ka,t(r,r′)−ka,t(r,Za,t)ha,t(r′)+ha,t(r)⊤Sa,tha,t(r′).(7)\Sigma_{d,a}(r,r',t)=k_{a,t}(r,r')-k_{a,t}(r,Z_{a,t})h_{a,t}(r')+h_{a,t}(r)^\top S_{a,t}h_{a,t}(r'). \tag*{(7)}

The vector mean μd(r,t)∈Rnd\mu_d(r,t)\in\mathbb{R}^{n_d} consists of the component means μd,a(r,t)∈R\mu_{d,a}(r,t)\in\mathbb{R}, while the cross-covariance Σd(r,r′,t)∈Rnd×nd\Sigma_d(r,r',t)\in\mathbb{R}^{n_d\times n_d} has diagonal entries Σd,a(r,r′,t)∈R\Sigma_{d,a}(r,r',t)\in\mathbb{R} and zero cross-component entries. POAM updates inducing locations using pivoted Cholesky decomposition (PCD), incorporates new measurements through analytic variational updates, and adapts kernel hyperparameters through mini-batch optimization. For further details, see Chen et al. [4].

Model-Based Diffusion

Model-Based Diffusion (MBD) [16] is a zeroth-order, sampling-based optimization method that iteratively refines candidate solutions through score-based reverse diffusion using samples, without requiring objective gradients.

Let θ∈Θ⊂Rnθ\theta\in\Theta\subset\mathbb{R}^{n_\theta}, with nθ∈N>0n_\theta\in\mathbb{N}_{>0}, denote the decision vector and J:Θ→RJ : \Theta\to\mathbb{R} its cost. MBD defines p0(θ)∝exp⁡(−J(θ)/λ)p_0(\theta) \propto\exp(-J(\theta)/\lambda) on Θ\Theta and p0=0p_0 = 0 outside Θ\Theta, with temperature λ∈R>0\lambda\in\mathbb{R}_{>0}. For ND∈N>0N_D \in\mathbb{N}_{>0} diffusion steps, indexed by ν=1,…,ND\nu= 1, \ldots, N_D, define αν∈(0,1)\alpha_\nu\in(0,1) and αˉν=∏k=1ναk\bar{\alpha}_\nu= \prod_{k=1}^{\nu} \alpha_k. At each step, Y(ν)∈RnθY^{(\nu)} \in\mathbb{R}^{n_\theta} denotes a noisy version of the decision vector θ\theta, generated by the forward diffusion process:

Y(ν)∣θ∼N(αˉνθ,(1−αˉν)Inθ).(8)Y^{(\nu)} \mid\theta\sim\mathcal{N}\left(\sqrt{\bar{\alpha}_\nu}\theta, (1-\bar{\alpha}_\nu)I_{n_\theta}\right). \tag*{(8)}

During reverse diffusion, Kθ∈N>0K_\theta\in\mathbb{N}_{>0} proposals are sampled and weighted by their costs:

θ(b)∼N(Y(ν)αˉν,1−αˉναˉνInθ),b=1,…,Kθ,θ^ν=∑b=1Kθθ(b)p0(θ(b))∑b=1Kθp0(θ(b)).(9)\theta^{(b)} \sim\mathcal{N}\left(\frac{Y^{(\nu)}}{\sqrt{\bar{\alpha}_\nu}}, \frac{1-\bar{\alpha}_\nu}{\bar{\alpha}_\nu}I_{n_\theta}\right), \quad b=1,\ldots,K_\theta,\qquad\hat{\theta}_\nu= \frac{\sum_{b=1}^{K_\theta}\theta^{(b)}p_0(\theta^{(b)})}{\sum_{b=1}^{K_\theta}p_0(\theta^{(b)})}. \tag*{(9)}

For a positive denominator, θ^ν∈Rnθ\hat{\theta}_\nu\in\mathbb{R}^{n_\theta} estimates the clean decision conditioned on Y(ν)Y^{(\nu)}. Let pνp_\nu be the marginal density of Y(ν)Y^{(\nu)} for θ∼p0\theta\sim p_0. Its score ∇log⁡pν(Y(ν))\nabla\log p_\nu(Y^{(\nu)}) is approximated by s^ν∈Rnθ\hat{s}_\nu\in\mathbb{R}^{n_\theta} in the reverse update:

s^ν=−Y(ν)+αˉνθ^ν1−αˉν,Y(ν−1)=Y(ν)+(1−αˉν)s^ναν.(10)\hat{s}_\nu= \frac{-Y^{(\nu)}+\sqrt{\bar{\alpha}_\nu}\hat{\theta}_\nu}{1-\bar{\alpha}_\nu},\qquad Y^{(\nu-1)} = \frac{Y^{(\nu)}+(1-\bar{\alpha}_\nu)\hat{s}_\nu}{\sqrt{\alpha_\nu}}. \tag*{(10)}

Repeating this update for ν=ND,…,1\nu=N_D,\ldots,1 progressively refines the candidates toward lower-cost solutions. Further details are provided by Pan et al. [16].

Problem Statement

Consider a robot governed by (1) monitoring (2), with initial state x0∈Sx_0 \in\mathcal{S}, clarity q0∈(0,1]Npq_0 \in(0,1]^{N_p}, and disturbance belief B0d=Bt0d\mathcal{B}_0^d = \mathcal{B}_{t_0}^d at t0t_0. A feedback policy π:X×(0,1]Np×Bd→U\pi: \mathcal{X} \times(0,1]^{N_p} \times\mathcal{B}_d \to\mathcal{U} selects u=π(x,q,Bd)u=\pi(x,q,\mathcal{B}^d) without access to the true field or future measurements. We seek to minimize the expected long-run mean clarity deficit subject to state and input constraints:

min⁡πlim sup⁡T→∞1TE[∫t0t0+Tqd(q(τ)) dτ],s.t.x˙=f(x,u,dˉ(Ψ(x))),q˙p=gp(x,qp),p∈Pcells,u=π(x,q,Btd)∈U,x∈S,t≥t0,x(t0)=x0,q(t0)=q0.(11)\begin{aligned} \min_{\pi}\quad&\limsup_{T\to\infty}\frac{1}{T}\mathbb{E}\left[\int_{t_0}^{t_0+T}q_d(q(\tau))\,\mathrm{d}\tau\right],\\ \text{s.t.}\quad&\dot{x}=f\left(x,u,\bar{d}(\Psi(x))\right),\\ &\dot{q}_p=g_p(x,q_p),\quad p\in\mathcal{P}_{\mathrm{cells}},\\ &u=\pi(x,q,\mathcal{B}_t^d)\in\mathcal{U},\quad x\in\mathcal{S},\quad t\ge t_0,\\ &x(t_0)=x_0,\quad q(t_0)=q_0. \tag*{(11)} \end{aligned}

Here T∈R>0T\in\mathbb{R}_{>0} is the averaging duration, not a terminal mission time. The expectation includes the initial disturbance uncertainty and observation noise; constraints are required almost surely for all t≥t0t\ge t_0. The belief Btd\mathcal{B}_t^d is updated from observations (5).

Directly solving (11) is computationally intractable, as it requires optimizing robot trajectories while accounting for future disturbance measurements and belief updates over an infinite horizon. We therefore consider the following problem and propose a safety-aware approximate solution to it.

Problem 1 (Disturbance-aware informative planning). Given the robot and monitoring models, environment geometry, and local disturbance measurements (5), develop an online method to approximate the constrained monitoring problem (11) by (i) learning the unknown disturbance field in regions relevant to the monitoring task and (ii) incorporating the learned disturbance belief into informative trajectory optimization while accounting for state and input constraints.

Methodology

Overview

We develop the two-stage online approach to Problem 1 illustrated in Figure 1. The first stage collects disturbance measurements in regions relevant to the monitoring task. Once the uncertainty in these regions is sufficiently reduced, the second stage uses the learned disturbance belief to optimize informative trajectories. Measurements of both the quantities of interest and the disturbance continue throughout both stages.

Figure 1. Overview of monitoring-informed disturbance estimation and disturbance-aware informative planning.

Let tit_i, i∈Ni \in\mathbb{N}, denote successive planning times, with ti+1>tit_{i+1} > t_i and ti→∞t_i \to\infty. We define xi=x(ti)∈Xx_i = x(t_i) \in\mathcal{X}, qi=[qp(ti)]p=1Np∈(0,1]Npq_i = [q_p(t_i)]_{p=1}^{N_p} \in(0,1]^{N_p}, and Bi=Btid\mathcal{B}_i = \mathcal{B}_{t_i}^{d} as the current robot state, clarity vector, and disturbance belief, respectively. The clarity targets are qˉ=[qˉp]p=1Np∈[0,1]Np\bar{q} = [\bar{q}_p]_{p=1}^{N_p} \in[0,1]^{N_p}. Trajectories are planned over a horizon TN∈R>0T_N \in\mathbb{R}_{>0} and executed in a receding-horizon manner. The planning index ii continues across both stages.

Geometric Path Initialization

We first generate geometric paths toward informative sensing locations using the current clarity map and known obstacle geometry. These paths identify regions where disturbance estimation can benefit the monitoring task and provide distinct initial trajectories for subsequent optimization.

Since the quantities of interest evolve over time, selecting sensing locations based only on the current clarity may overlook regions with significant future clarity deficits. We therefore predict clarity over [ti,ti+TN][t_i,t_i+T_N], assuming no additional measurements are collected. Setting Cp=0C_p = 0 in (3), the predicted clarity qi,pns(τ)q_{i,p}^{\mathrm{ns}}(\tau) of cell p∈Pcellsp \in\mathcal{P}_{\mathrm{cells}} at time ti+τt_i+\tau, for τ∈[0,TN]\tau\in[0,T_N], and its corresponding deficit are

qi,pns(τ)=qp(ti)1+Qpqp(ti)τ,δi,pns(τ)=[qˉp−qi,pns(τ)]+.(12)q_{i,p}^{\mathrm{ns}}(\tau) = \frac{q_p(t_i)}{1+Q_p q_p(t_i)\tau}, \qquad\delta_{i,p}^{\mathrm{ns}}(\tau) = [\bar{q}_p-q_{i,p}^{\mathrm{ns}}(\tau)]_{+}. \tag*{(12)}

We evaluate these deficits at Nτ∈N>0N_{\tau} \in\mathbb{N}_{>0} selected times τh∈[0,TN]\tau_h \in[0,T_N], h=1,…,Nτh = 1,\ldots,N_{\tau}, forming Dins=[δi,pns(τh)]p,h∈[0,1]Np×NτD_i^{\mathrm{ns}} = [\delta_{i,p}^{\mathrm{ns}}(\tau_h)]_{p,h} \in[0,1]^{N_p \times N_{\tau}}. This accounts for how clarity deficits develop throughout the prediction horizon.

Using these predictions and known obstacle geometry, we generate a fixed number Nγ∈N>0N_{\gamma} \in\mathbb{N}_{>0} of alternative collision-free paths. Let Pfree⊂P\mathcal{P}_{\mathrm{free}} \subset\mathcal{P} denote the collision-free workspace and Γi={γiℓ}ℓ=1Nγ\Gamma_i = \{\gamma_i^{\ell}\}_{\ell=1}^{N_{\gamma}} the generated paths, where ℓ=1,…,Nγ\ell= 1,\ldots,N_{\gamma} indexes individual paths. Each path γiℓ:[0,1]→Pfree\gamma_i^{\ell} : [0,1] \to\mathcal{P}_{\mathrm{free}} starts at Ψ(xi)\Psi(x_i) and reaches a candidate sensing location, with s∈[0,1]s \in[0,1] denoting normalized arc length. Alternative paths through different passages are retained to provide geometrically distinct routes.

To prioritize paths according to their monitoring relevance, we evaluate the predicted clarity deficits of cells observable along each path. Let xiℓ(s)∈Xx_i^{\ell}(s) \in\mathcal{X} be a sensing configuration at γiℓ(s)\gamma_i^{\ell}(s), satisfying Ψ(xiℓ(s))=γiℓ(s)\Psi(x_i^{\ell}(s)) = \gamma_i^{\ell}(s). The score of the ℓ\ell-th path is defined as

aiℓ=1Nτ∑h=1Nτ∫01∑p=1NpCp(xiℓ(s))2δi,pns(τh) ds.(13)a_i^{\ell} = \frac{1}{N_{\tau}} \sum_{h=1}^{N_{\tau}} \int_{0}^{1} \sum_{p=1}^{N_p} C_p(x_i^{\ell}(s))^2 \delta_{i,p}^{\mathrm{ns}}(\tau_h)\,\mathrm{d}s. \tag*{(13)}

The integral can be approximated numerically by evaluating the integrand at discrete points along each path. A larger score indicates that the path passes through locations from which cells with greater predicted clarity deficits can be observed. These scores form ai=[aiℓ]ℓ=1Nγ∈R≥0Nγa_i = [a_i^{\ell}]_{\ell=1}^{N_{\gamma}} \in\mathbb{R}_{\geq0}^{N_{\gamma}} and are subsequently used to prioritize disturbance estimation along the geometric paths. The generated paths are also used to initialize the subsequent MBD trajectory optimization.

Disturbance Estimation

We use the geometric paths and their monitoring relevance scores from § 4.2 to determine where disturbance measurements should be prioritized.

Monitoring-informed spatial weighting. Disturbances along geometric paths can affect the robot’s motion and its ability to collect measurements of the quantities of interest. We therefore construct spatial weights that assign greater importance to regions near paths with higher monitoring relevance. Let κρ(r,γ)∈R≥0\kappa_{\rho}(r,\gamma) \in\mathbb{R}_{\geq0} be a prescribed spatial weighting function that assigns greater weight to locations near the geometric path γ\gamma, where r∈Pfreer \in\mathcal{P}_{\mathrm{free}} and ρ∈R>0\rho\in\mathbb{R}_{>0} is a tunable parameter. For each cell p∈Pcellsp \in\mathcal{P}_{\mathrm{cells}}, define an unnormalized weight and, when their sum is positive, its normalized value by

ω~i,p=∑ℓ=1Nγaiℓκρ(rp,γiℓ),ωi,p=ω~i,pWi,(14)\widetilde{\omega}_{i,p} = \sum_{\ell=1}^{N_{\gamma}} a_i^{\ell}\kappa_{\rho}(r_p,\gamma_i^{\ell}), \qquad\omega_{i,p} = \frac{\widetilde{\omega}_{i,p}}{W_i}, \tag*{(14)}

where Wi=∑p=1Npω~i,pW_i = \sum_{p=1}^{N_p} \widetilde{\omega}_{i,p} is the normalization factor. For Wi>0W_i > 0, the weights ωi=[ωi,p]p=1Np∈R≥0Np\omega_i = [\omega_{i,p}]_{p=1}^{N_p} \in\mathbb{R}_{\geq0}^{N_p} sum to one.

Disturbance measurement selection. Using the spatial weights, we quantify the disturbance uncertainty in regions relevant to the monitoring task as

Ui=(∑p=1Npωi,ptr⁡Pd(rp,ti))1/2.(15)U_i = \left(\sum_{p=1}^{N_p} \omega_{i,p}\operatorname{tr} P_d(r_p,t_i)\right)^{1/2}. \tag*{(15)}

We aim to select measurements that reduce this uncertainty. Since the disturbance belief captures spatial correlations, a measurement at one location can reduce uncertainty at other locations. For a candidate measurement location z∈Pfreez \in\mathcal{P}_{\mathrm{free}}, let Gi(z)∈R≥0G_i(z) \in\mathbb{R}_{\geq0} denote the expected reduction in Ui2U_i^2 from a disturbance measurement at zz. Using Gaussian conditioning, this reduction is given by

Gi(z)=∑p=1Npωi,ptr⁡[Σd(rp,z,ti)(Pd(z,ti)+Rd)−1Σd(z,rp,ti)].(16)G_i(z) = \sum_{p=1}^{N_p} \omega_{i,p} \operatorname{tr}\left[\Sigma_d(r_p,z,t_i)\left(P_d(z,t_i)+R_d\right)^{-1}\Sigma_d(z,r_p,t_i)\right]. \tag*{(16)}

This one-measurement reduction is evaluated with the current spatial weights, GP kernel parameters, and inducing representation held fixed; subsequent POAM retraining is not predicted. To account for travel cost, we consider a finite set of candidate measurement locations Zi⊂Pfree\mathcal{Z}_i \subset\mathcal{P}_{\mathrm{free}}. Let Li:Zi→R≥0L_i:\mathcal{Z}_i \to\mathbb{R}_{\geq0} give the collision-free travel distance from Ψ(xi)\Psi(x_i) to each candidate, and let L0∈R>0L_0 \in\mathbb{R}_{>0} control the travel-distance penalty. When candidates remain after route checks, we select

zi⋆∈arg⁡max⁡z∈ZiGi(z)1+Li(z)/L0.(17)z_i^\star\in\arg\max_{z \in\mathcal{Z}_i} \frac{G_i(z)}{1+L_i(z)/L_0}. \tag*{(17)}

We exclude candidate routes that intersect obstacles.

Stopping criterion for disturbance estimation. At each planning iteration, UiU_i is updated using the current clarity and disturbance belief. Dedicated disturbance estimation terminates when Ui≤εdU_i \leq\varepsilon_d, where εd∈R>0\varepsilon_d \in\mathbb{R}_{>0} is a prescribed uncertainty tolerance. Throughout this stage, clarity continues to evolve, and the monitoring cost incurred during disturbance estimation is included in the evaluation.

Disturbance-Aware Model-Based Diffusion

In the second stage, we incorporate the learned disturbance belief into Model-Based Diffusion (MBD) to optimize informative trajectories. At each planning time tit_i, geometric paths from § 4.2 initialize candidate trajectories, which are refined through MBD. Each candidate is evaluated by predicting the robot’s motion under sampled disturbance fields and the resulting information acquisition, allowing the optimizer to account for disturbance effects directly.

Disturbance-aware trajectory prediction. Let θ∈Θ⊂Rnθ\theta\in\Theta\subset\mathbb{R}^{n_\theta} parameterize candidate state and input curves ξˉ(⋅;θ)=(ξˉx,ξˉu)\bar{\xi}(\cdot;\theta) = (\bar{\xi}_x,\bar{\xi}_u) over [ti,ti+TN][t_i,t_i+T_N]. To evaluate each candidate, we draw Nd∈N>0N_d \in\mathbb{N}_{>0} spatial disturbance fields di(s):P→Rndd_i^{(s)}:\mathcal{P}\to\mathbb{R}^{n_d} from Bid\mathcal{B}_i^d, indexed by s=1,…,Nds=1,\ldots,N_d. Each sampled field remains fixed throughout its rollout, preserving spatial correlations.

Let Ti={ξˉ(⋅;θ):θ∈Θ}\mathcal{T}_i=\{\bar{\xi}(\cdot;\theta):\theta\in\Theta\} denote the set of candidate trajectories. For each sample, the tracking controller πtr:[ti,ti+TN]×X×Ti×Bd→Rm\pi_{\mathrm{tr}}:[t_i,t_i+T_N]\times\mathcal{X}\times\mathcal{T}_i\times\mathcal{B}_d\to\mathbb{R}^m computes a nominal control input to follow ξˉ(⋅;θ)\bar{\xi}(\cdot;\theta) using the current predicted state and disturbance belief. This input is modified by an approximate safety filter Fpred:X×Rm×Bd→U\mathcal{F}^{\mathrm{pred}}:\mathcal{X}\times\mathbb{R}^m\times\mathcal{B}_d\to\mathcal{U}, based on the approach described in the following paragraph. The resulting predicted state ξxθ,s(t)∈Rn\xi_x^{\theta,s}(t)\in\mathbb{R}^n and input ξuθ,s(t)∈Rm\xi_u^{\theta,s}(t)\in\mathbb{R}^m satisfy

unomθ,s(t)=πtr(t,ξxθ,s(t),ξˉ(⋅;θ),Bid),u_{\mathrm{nom}}^{\theta,s}(t)=\pi_{\mathrm{tr}}\left(t,\xi_x^{\theta,s}(t),\bar{\xi}(\cdot;\theta),\mathcal{B}_i^d\right),
ξuθ,s(t)=Fpred(ξxθ,s(t),unomθ,s(t),Bid),\xi_u^{\theta,s}(t)=\mathcal{F}^{\mathrm{pred}}\left(\xi_x^{\theta,s}(t),u_{\mathrm{nom}}^{\theta,s}(t),\mathcal{B}_i^d\right),
ξ˙xθ,s(t)=f(ξxθ,s(t),ξuθ,s(t),di(s)(Ψ(ξxθ,s(t)))).(18)\dot{\xi}_x^{\theta,s}(t)=f\left(\xi_x^{\theta,s}(t),\xi_u^{\theta,s}(t),d_i^{(s)}\left(\Psi\left(\xi_x^{\theta,s}(t)\right)\right)\right). \tag*{(18)}

with ξxθ,s(ti)=xi\xi_x^{\theta,s}(t_i)=x_i. The clarity vector qθ,s(t)∈(0,1]Npq^{\theta,s}(t)\in(0,1]^{N_p} is propagated using (3), initialized at qiq_i. Thus, disturbances and filter interventions affect the predicted information acquisition.

Safety-aware filtering during planning. We now describe the approximate safety filter Fpred\mathcal{F}^{\mathrm{pred}} used in (18), based on the robust policy control barrier function (RPCBF) approach [8]. Let h:X→Rh:\mathcal{X}\to\mathbb{R} encode the admissible state set as S={x∈X:h(x)≤0}\mathcal{S}=\{x\in\mathcal{X}:h(x)\le0\}, and let πb:X→U\pi_b:\mathcal{X}\to\mathcal{U} denote a prescribed design policy satisfying the input constraints. At each predicted state x=ξ^xθ,s(t)x=\widehat{\xi}^{\theta,s}_x(t) along a candidate trajectory, the RPCBF estimates the maximum future constraint violation by simulating πb\pi_b under the sampled disturbance fields di(s)d_i^{(s)}. The sampled finite-horizon value function is

V^i(x)=max⁡s′=1,…,Ndsup⁡0≤τ≤TNh(xiπb,s′(τ;x))(19)\widehat{V}_i(x)=\max_{s'=1,\ldots,N_d}\sup_{0\le\tau\le T_N}h\left(x_i^{\pi_b,s'}(\tau;x)\right) \tag*{(19)}

where xiπb,s′(τ;x)∈Xx_i^{\pi_b,s'}(\tau;x)\in\mathcal{X} denotes the auxiliary state under πb\pi_b and field di(s′)d_i^{(s')}, initialized at xiπb,s′(0;x)=xx_i^{\pi_b,s'}(0;x)=x. The inner index s′s' ranges over all NdN_d fields while the outer rollout index ss is fixed. The sample set is held fixed when estimating the gradient with respect to xx. Using V^i(x)\widehat{V}_i(x) and its estimated gradient, the filter modifies the nominal tracking input unomθ,s(t)u_{\mathrm{nom}}^{\theta,s}(t) by minimizing its deviation subject to the RPCBF constraint and input bounds u∈Uu\in\mathcal{U}. For control-affine dynamics and polyhedral input bounds, imposing the sampled barrier conditions yields a quadratic program (QP) [8]. The filtered input is used to propagate the predicted robot motion in (18), with Fpred\mathcal{F}^{\mathrm{pred}} applied throughout each candidate rollout. The purpose of this filter is to account for how potential safety interventions affect the predicted motion and information acquisition, rather than to guarantee safety during planning.

Trajectory cost and optimization. Each candidate is evaluated using the accumulated clarity deficit along its predicted motion, with a soft penalty for state-constraint violations. Using the constraint function hh defined above and a penalty coefficient η∈R≥0\eta\in\mathbb{R}_{\ge0}, the finite-horizon cost is

J^i(θ)=1Nd∑s=1Nd∫titi+TN[qd(qθ,s(t))+η[h(ξ^xθ,s(t))]2] dt.(20)\widehat{J}_i(\theta)=\frac{1}{N_d}\sum_{s=1}^{N_d}\int_{t_i}^{t_i+T_N}\left[q_d\left(q^{\theta,s}(t)\right)+\eta\left[h\left(\widehat{\xi}^{\theta,s}_x(t)\right)\right]^2\right]\,dt. \tag*{(20)}

The cost accounts for disturbance effects and safety-filter interventions through the predicted state and clarity trajectories. The geometric paths from § 4.2 initialize separate MBD optimizations using (8)–(10), with J=J^iJ=\widehat{J}_i. Each optimization refines its candidate, and the lowest-cost evaluated candidate with a finite cost is selected, with parameters θi∗∈Θ\theta_i^*\in\Theta.

Execution and belief updates. The robot executes an initial segment of the selected candidate using the tracking controller and runtime filter Fexec\mathcal{F}^{\mathrm{exec}}, which has the same argument and output spaces as Fpred\mathcal{F}^{\mathrm{pred}}, then replans from the updated state and clarity. Actual disturbance measurements continue updating BidB_i^d through POAM during both stages.

Results and Discussion

We evaluate four environments with spatially varying, time-invariant wind, using double-integrator nominal dynamics and a relative-air-velocity force model. Methods share each scenario’s geometry, initial state and clarity, and MBD budget. The planning horizon is 15 s, with 5 s execution intervals and no initial wind measurements. Appendix § 8 provides simulation details.

Baselines and evaluation: We compare B1, zero-wind MBD without filtering; B2, zero-wind MBD with the learned wind belief used in tracking and runtime RPCBF; B3, disturbance-aware MBD with RPCBF during planning and execution, learning only from measurements collected while monitoring; and B4, global wind learning followed by disturbance-aware MBD. The proposed method targets learning along monitoring-relevant paths before using the same planner as B4.

Figure 2 compares monitoring histories, while Table 1 in Appendix § 8 reports the final mean clarity deficit D(T)D(T) and its time average DT=T−1∫0TD(t) dtD_T = T^{-1}\int_{0}^{T}D(t)\,\mathrm{d}t, with D(t)=qd(q(t))D(t) = q_d(q(t)) and elapsed time measured from the start of each run. Contact terminates a run, and monitoring deficits are not extrapolated beyond that point.

Figure 2. Monitoring across four scenarios. Top: environments and proposed trajectories; bottom: mean clarity-deficit histories for all methods. Comparison durations are 532, 560, 544, and 514 s for Scenarios 1–4, respectively.

MethodScenario 1Scenario 2Scenario 3Scenario 4
T=532T=532 sT=560T=560 sT=544T=544 sT=514T=514 s
B1Contact: 4.4 sContact: 15.6 sContact: 4.8 s0.181 (0.161)
B20.293 (0.366)0.301 (0.367)0.078 (0.081)0.180 (0.180)
B30.224 (0.280)0.221 (0.267)0.047 (0.035)0.186 (0.184)
B4Contact: 399.0 s0.190 (0.156)0.054 (0.040)0.221 (0.177)
Proposed0.166 (0.146)0.177 (0.171)0.042 (0.027)0.189 (0.189)

Table 1. Monitoring performance: DTD_T, with final deficit D(T)D(T) in parentheses. Bold indicates the lowest DTD_T among completed runs; contact times are reported for terminated runs.

Route-dependent disturbances: In Scenario 1, wind opposes travel from region B to A through the central passage, while a sheltered lower bypass is available. B2 and B3 do not enter A within the comparison window, whereas the proposed method does, reducing DTD_T from B3’s 0.224 to 0.166. B4 also reaches A but contacts an obstacle at 399.0 s. In Scenario 2, wind also opposes travel through the lower passage, making the upper bypass important. Again, B2 and B3 do not enter A. Targeted learning ends at 200 s versus B4’s 370 s and yields lower DTD_T, although B4 has the lower final deficit. The improvement therefore concerns accumulated monitoring performance, not every instantaneous deficit.

Spatial allocation of learning: In Scenario 3, the proposed method learns disturbances along paths connecting the two monitoring regions rather than throughout the workspace. By focusing learning on monitoring-relevant paths, the proposed method reduces estimation time from 244 to 124 s and lowers DTD_T from 0.0540 to 0.0418 (22.6%) relative to B4. In Scenario 4, the monitoring region is largely calm. Although the proposed method reduces learning time from 214 to 102 s relative to B4, B2 achieves a lower DTD_T, showing that dedicated learning may be unnecessary when disturbances have little effect on monitoring.

Conclusion

We presented a two-stage framework combining monitoring-informed disturbance learning with disturbance-aware trajectory optimization. In the four reported runs, monitoring-informed sampling reduces dedicated learning time by 25.0–52.3% relative to global estimation and improves monitoring when disturbances restrict access to sensing locations. The largely calm scenario also exposes the cost of unnecessary learning.

Algorithmic Details

This appendix details Algorithm 1, which switches from disturbance estimation to persistent informative planning when the weighted disturbance uncertainty satisfies Ui≤ϵdU_i \le\epsilon_d.

Planning state and geometric initialization

At planning time tit_i, the available quantities are xi∈Xx_i \in\mathcal{X}, qi∈(0,1]Npq_i \in(0,1]^{N_p}, the disturbance data Dtid\mathcal{D}^{d}_{t_i}, and their belief Bid\mathcal{B}^{d}_{i}. The geometry, target clarity qˉ\bar{q}, horizon TNT_N, uncertainty tolerance ϵd\epsilon_d, and remaining design parameters are fixed inputs. The index i∈Ni \in\mathbb{N} counts planning iterations and is not reset between stages. The retained waypoint is z∈Pfreez \in\mathcal{P}_{\mathrm{free}}, or z=∅z = \varnothing when none is available. The symbol ξˉi\bar{\xi}_i denotes the candidate selected for execution at tit_i, or ∅\varnothing when no candidate is available.

The operation GeoPaths(xi,qi)(x_i,q_i) predicts clarity deficits using (12) and constructs geometrically distinct paths Γi\Gamma_i, scored over their entire length by (13). Numerical quadrature approximates the path integrals. These paths guide disturbance learning and initialize separate MBD optimizations; neither their construction nor scoring requires the disturbance belief.

The operation PathWeights(Γi,ai)(\Gamma_i,a_i) uses κρ\kappa_{\rho} in (14) to distribute path scores to nearby cells and normalize the weights ωi\omega_i. It also returns their unnormalized sum WiW_i. If Wi=0W_i = 0, the algorithm sets ωi=0\omega_i = 0 and Ui=+∞U_i = +\infty, preventing an undefined normalization or premature transition to monitoring.

Disturbance measurement selection

For positive total weight, the algorithm evaluates UiU_i using (15). If Ui≤ϵdU_i \le\epsilon_d, it leaves the first loop before selecting another disturbance-measurement waypoint. Otherwise, SelectWaypoint retains a usable current waypoint or applies (17) to the candidate set ZiZ_i. A new waypoint is needed after arrival, rejection of the connecting route, or lack of progress. Candidates are excluded when their routes intersect obstacles or are predicted to require inputs outside U\mathcal{U}. If Wi=0W_i = 0 or no usable candidate remains, the routine returns ∅\varnothing.

The acquisition value Gi(z)G_i(z) in (16) predicts the reduction in weighted disturbance variance from one measurement, holding spatial weights and GP parameters fixed; subsequent POAM updates are not predicted. The travel penalty in (17) accounts for distance, not the full downstream monitoring benefit. Measurements acquired during execution update the disturbance belief.

Given a waypoint, MeasurementPath(ti,xi,z,Bid)(t_i,x_i,z,\mathcal{B}^{d}_{i}) constructs a candidate trajectory ξˉi\bar{\xi}_i toward it using the known geometry and current disturbance belief. It returns ∅\varnothing if z=∅z = \varnothing or no usable trajectory is available. A prescribed recovery policy is then used during execution. The specific recovery controller is an implementation choice, not an additional safety guarantee.

Disturbance-aware MBD optimization

After the transition, Initialize(Γi,xi)(\Gamma_i, x_i) time-parameterizes the geometric paths to form a finite set Θi0⊂Θ\Theta_i^0 \subset\Theta of candidate parameter vectors. Available hold and shifted previous candidates may also be retained. Each initialization is refined independently using the MBD equations in § 2.6; candidate parameters from different chains are not averaged together.

The operation MBD(ti,xi,qi,Bid,Θi0)(t_i, x_i, q_i, \mathcal{B}_i^d, \Theta_i^0) evaluates each candidate using the sampled fields di(s)d_i^{(s)}, the filtered dynamics (18), and the cost (20). A sampled field is held fixed along its complete rollout. At each predicted state, the tracking input is modified by Fpred\mathcal{F}^{\mathrm{pred}}, whose RPCBF value estimate is constructed under the design policy as in (19). These design-policy predictions are auxiliary calculations for the filter; the candidate itself is propagated using the filtered tracking input. Clarity is propagated along the resulting state trajectory, and the disturbance belief remains fixed within the planning call.

The cost is averaged over disturbance samples before computing the MBD proposal weights. Initial candidates, sampled proposals, and refined means are evaluated before selection. If all proposal weights vanish, no weighted update is defined and that chain retains its previously evaluated candidates. Among candidates with defined, finite evaluations, the routine selects the lowest-cost trajectory and returns its parameters θi∗∈Θ\theta_i^* \in\Theta; if none is available, it returns ∅\varnothing. A finite set of predicted trajectories does not certify physical safety. No hypothetical measurement is added to Did\mathcal{D}_i^d during this optimization.

Execution and actual-data updates

The operation Executei(ξˉi)_i(\bar{\xi}_i) follows an initial segment of the candidate using the tracking controller and runtime filter Fexec\mathcal{F}^{\mathrm{exec}}. If ξˉi=∅\bar{\xi}_i = \varnothing, the prescribed recovery controller supplies an input in U\mathcal{U}. In the first stage the candidate leads toward a disturbance-measurement waypoint; in the second it is selected by MBD. The routine returns the next planning time ti+1t_{i+1}, state xi+1x_{i+1}, clarity qi+1q_{i+1}, and new disturbance observations

ΔDid={(rj,yjd):ti<tjd≤ti+1}.\Delta\mathcal{D}_i^d = \left\{(r_j, y_j^d) : t_i < t_j^d \le t_{i+1}\right\}.

The operation Ud\mathcal{U}_d incorporates the accumulated observations into the POAM model, including its online variational and kernel updates. The dataset union in Algorithm 1 records this assimilation and does not imply that the model must be retrained from scratch. Measurements of the quantities of interest also remain active, so clarity evolves throughout both stages. The planning clock, clarity, and disturbance belief are carried across the transition without resetting. If the uncertainty threshold is never met, the first stage continues; the algorithm assumes no terminal mission time.

Algorithm 1: Persistent disturbance-aware informative planning. The estimation stage ends only when weighted posterior uncertainty meets its tolerance.

Input: t0,x0,q0,B0d,D0dt_0, x_0, q_0, \mathcal{B}_0^d, \mathcal{D}_0^d; models, geometry, qˉ\bar{q}, and design parameters

  1. i←0,z←∅i \leftarrow0,\quad z \leftarrow\emptyset

  2. // Stage 1: Disturbance estimation

  3. while true do

  4. (Γi,ai)←GeoPaths⁡(xi,qi)\left(\Gamma_i, a_i\right) \leftarrow\operatorname{GeoPaths}(x_i, q_i)

  5. (ωi,Wi)←PathWeights⁡(Γi,ai)\left(\omega_i, W_i\right) \leftarrow\operatorname{PathWeights}(\Gamma_i, a_i)

  6. Ui←+∞U_i \leftarrow+\infty

  7. if Wi>0W_i > 0 then Ui←(∑p=1Npωi,ptr⁡Pd(rp,ti))1/2U_i \leftarrow\left(\sum_{p=1}^{N_p}\omega_{i,p}\operatorname{tr} P_d(r_p,t_i)\right)^{1/2}

  8. if Ui≤εdU_i \le\varepsilon_d then break

  9. z←SelectWaypoint⁡(z,xi,ωi,Bid)z \leftarrow\operatorname{SelectWaypoint}(z, x_i, \omega_i, \mathcal{B}_i^d)

  10. ξˉi←MeasurementPath⁡(ti,xi,z,Bid)\bar{\xi}_i \leftarrow\operatorname{MeasurementPath}(t_i, x_i, z, \mathcal{B}_i^d)

  11. (ti+1,xi+1,qi+1,ΔDid)←Execute⁡i(ξˉi)\left(t_{i+1},x_{i+1},q_{i+1},\Delta\mathcal{D}_i^d\right) \leftarrow\operatorname{Execute}_i(\bar{\xi}_i)

  12. Di+1d←Did∪ΔDid\mathcal{D}_{i+1}^d \leftarrow\mathcal{D}_i^d \cup\Delta\mathcal{D}_i^d

  13. Bi+1d←Ud(Di+1d)\mathcal{B}_{i+1}^d \leftarrow\mathcal{U}_d(\mathcal{D}_{i+1}^d)

  14. i←i+1i \leftarrow i+1

  15. // Stage 2: Disturbance-aware informative planning

  16. while true do

  17. (Γi,ai)←GeoPaths⁡(xi,qi)\left(\Gamma_i, a_i\right) \leftarrow\operatorname{GeoPaths}(x_i, q_i)

  18. Θi0←Initialize⁡(Γi,xi)\Theta_i^0 \leftarrow\operatorname{Initialize}(\Gamma_i, x_i)

  19. θi∗←MBD⁡(ti,xi,qi,Bid,Θi0)\theta_i^\ast\leftarrow\operatorname{MBD}(t_i, x_i, q_i, \mathcal{B}_i^d, \Theta_i^0)

  20. ξˉi←∅\bar{\xi}_i \leftarrow\emptyset

  21. if θi∗≠∅\theta_i^\ast\ne\emptyset then ξˉi←ξˉ(⋅;θi∗)\bar{\xi}_i \leftarrow\bar{\xi}(\cdot;\theta_i^\ast)

  22. (ti+1,xi+1,qi+1,ΔDid)←Execute⁡i(ξˉi)\left(t_{i+1},x_{i+1},q_{i+1},\Delta\mathcal{D}_i^d\right) \leftarrow\operatorname{Execute}_i(\bar{\xi}_i)

  23. Di+1d←Did∪ΔDid\mathcal{D}_{i+1}^d \leftarrow\mathcal{D}_i^d \cup\Delta\mathcal{D}_i^d

  24. Bi+1d←Ud(Di+1d)\mathcal{B}_{i+1}^d \leftarrow\mathcal{U}_d(\mathcal{D}_{i+1}^d)

  25. i←i+1i \leftarrow i+1

Environments, Baselines, and Detailed Results

Four scenarios assess disturbance-aware trajectory prediction, alternative-route discovery, and monitoring-informed disturbance learning. Wind affects robot motion while the QoI dynamics remain separate. Figure 3, Figure 2, and Figure 4 show the environments and trajectory comparisons. Simulations were performed on an Apple M3 MacBook Air with 16 GB of unified memory.

Figure 3. The four evaluation environments. Scenarios 1 and 2 test alternative passages under opposing wind; Scenario 3 has two monitoring regions within a larger workspace; Scenario 4 has a largely calm monitoring region. The lower-left and lower-right panels correspond to Scenarios 3 and 4 in the main text; their embedded labels are reversed.

Figure 4. Trajectories of the proposed method and baselines, with rows corresponding to Scenarios 1–4 in the main text. The embedded labels in the final two rows are reversed. Runs ending in contact are terminated, not completed comparisons; Table 1 reports the numerical results.

Simulation and comparison protocol

Simulations use double-integrator nominal dynamics with wind entering through a relative-air-velocity force model. Wind varies spatially but remains fixed in time. Methods share each scenario’s geometry, initial state and clarity, and MBD budget. The planning horizon is 15 s, with 5 s execution intervals. Wind estimation begins without measurements, and POAM is updated online from observations collected during execution.

Dedicated learning terminates when the path-weighted disturbance uncertainty in (15) falls below εd=1.10\varepsilon_d = 1.10 m/s. B4 instead evaluates disturbance uncertainty throughout the free workspace. This posterior uncertainty criterion is model-dependent and does not bound the true wind-estimation error.

Scenarios 1–4 are evaluated over 532, 560, 544, and 514 s, respectively, measured from mission start and including dedicated learning. QoI sensing remains active during learning. These evaluation windows do not impose a terminal time on the persistent planner. Runs end upon contact, and monitoring deficits are not extrapolated thereafter.

Comparison methods

B1: MBD without RPCBF. B1 assumes zero wind in planning and executes MBD trajectories without safety filtering.

B2: MBD with runtime RPCBF. B2 uses zero-wind planning, while tracking and runtime RPCBF use the learned wind belief. The planner therefore does not account for disturbance-induced motion or filter corrections when choosing trajectories.

B3: Disturbance-aware MBD without dedicated learning. B3 begins informative monitoring immediately, evaluating candidate trajectories using the learned wind belief and RPCBF during planning and execution. POAM learns from measurements collected during monitoring, without a dedicated disturbance-learning stage.

B4: Global learning followed by disturbance-aware MBD. B4 reduces wind uncertainty throughout the free workspace, including regions irrelevant to monitoring, before using the same disturbance-aware MBD planner as the proposed method.

Proposed: Monitoring-informed learning followed by disturbance-aware MBD. The proposed method prioritizes disturbance measurements near monitoring-relevant geometric paths, including transit passages, before using the same disturbance-aware MBD planner as B4. Learning allocation is handled in the first stage rather than by adding a disturbance-information reward to the monitoring cost.

Metrics and learning durations

For each simulation, define the mean clarity deficit and its time average as

D(t)=1Np∑p=1Np[qˉp−qp(t)]+,DT=1T∫0TD(t) dt.(21)D(t)=\frac{1}{N_p}\sum_{p=1}^{N_p}\left[\bar{q}_p-q_p(t)\right]_+,\qquad D_T=\frac{1}{T}\int_0^T D(t)\,dt. \tag*{(21)}

The final deficit D(T)D(T) depends on the revisit cycle, whereas DTD_T captures monitoring loss over the full comparison window. Table 1 reports both; bold denotes the lowest DTD_T among completed runs, not a statistically significant difference.

Table 2 reports simulated disturbance-learning times and percentage reductions 100(tB4−tprop)/tB4100(t_{\mathrm{B4}}-t_{\mathrm{prop}})/t_{\mathrm{B4}}, where tB4t_{\mathrm{B4}} and tpropt_{\mathrm{prop}} are the respective durations. These are mission times, not computation times; QoI sensing remains active throughout learning.

ScenarioB4 (s)Proposed (s)Reduction
124018025.0%
237020045.9%
324412449.2%
421410252.3%

Table 2. Dedicated wind-estimation time.

Scenario 1: Central opposing wind and a lower bypass

Wind opposes travel from region B to A through the central passage, while a sheltered lower bypass remains available. B2 and B3 do not reach A within 532 s, although B3 achieves a lower mean deficit than B2. The proposed method reaches A and reduces DTD_T to 0.166 versus 0.224 for B3. B4 also reaches A but contacts an obstacle at 399.0 s; it therefore has no full-window monitoring score.

Scenario 2: Opposing wind in the central and lower passages

Wind also opposes the lower passage, making the upper bypass important. B2 and B3 again fail to reach A within the 560 s comparison window. Both dedicated-learning methods reach A. Targeted learning takes 200 s versus 370 s for B4. The proposed method reduces DTD_T from 0.190 to 0.177, although B4 has the lower final deficit (0.156 versus 0.171).

Scenario 3: Monitoring concentrated in two regions

Two monitoring regions are embedded in an obstacle-filled workspace with additional wind outside monitoring-relevant routes. B2 never reaches A, while B3 reaches it only at 351.55 s, illustrating the delay in discovering an alternative without dedicated disturbance learning. Compared with B4, targeted learning ends after 124 rather than 244 s, leaving more time for monitoring. The proposed method reduces DTD_T from 0.0540 to 0.0418 (22.6%) and D(T)D(T) from 0.0396 to 0.0269 (about 32%). B4 reaches A earlier during learning (54.10 versus 74.65 s), so the proposed method’s improvement reflects a shorter learning stage rather than an earlier first arrival.

Scenario 4: Calm monitoring region and remote wind

The monitoring region is largely calm, with stronger wind elsewhere. Targeted learning takes 102 s versus B4’s 214 s and lowers DTD_T from 0.221 to 0.189, although its final deficit is higher (0.189 versus 0.177). However, B2 attains the lowest DT=0.180D_T = 0.180 (B1: 0.181; B3: 0.186). Dedicated learning may therefore be unnecessary when disturbances have little effect on monitoring, and the fixed switching threshold still incurs learning cost in this scenario.

Planning time and interpretation

In Scenario 3, median planning times after compilation are 1.64 s (proposed) and 1.60 s (B4), with 95th percentiles of 1.93 and 1.73 s, respectively. Both methods use the same MBD budget. Timings include posterior and field preparation, geometric initialization, optimization, and safety-belief refresh. The baselines distinguish execution-time compensation (B2), disturbance-aware monitoring without dedicated learning (B3), and spatial allocation of learning with a common monitoring planner (B4 versus proposed). These comparisons do not separately isolate geometric initialization, weighting, or the switching rule. Each scenario reports one run. The proposed method completes all four comparison windows without contact, but infeasible RPCBF steps occur, and B4 contacts an obstacle in Scenario 1 despite filtering. The results do not establish guaranteed collision avoidance or statistical robustness. The path weights approximate task relevance rather than the exact long-term value of disturbance information.

References

  1. [1]Devansh R. Agrawal and Dimitra Panagou. Sensor-based planning and control for robotic systems: Introducing clarity and perceivability. IEEE Control Systems Letters, 7:2623–2628, 2023. doi: 10.1109/LCSYS.2023.3288493. URL https://arxiv.org/abs/2304.02578.
  2. [2]Elena Arcari, Lukas Hewing, Max Schlichting, and Melanie Zeilinger. Dual stochastic MPC for systems with parametric and structural uncertainty. In Proceedings of the 2nd Conference on Learning for Dynamics and Control, volume 120 of Proceedings of Machine Learning Research, pages 894–903, 2020. URL https://proceedings.mlr.press/v120/arcari20a.html.
  3. [3]Jonathan Binney, Andreas Krause, and Gaurav S. Sukhatme. Optimizing waypoints for monitoring spatiotemporal phenomena. The International Journal of Robotics Research, 32(8):873–888, 2013. doi: 10.1177/0278364913488427.DOI
  4. [4]Weizhe Chen, Lantao Liu, and Roni Khardon. POAM: Probabilistic online attentive mapping for efficient robotic information gathering. In Robotics: Science and Systems, 2024. doi: 10.15607/RSS.2024.XX.019.DOI
  5. [5]Doo-Hyun Cho, Jung-Su Ha, Sujin Lee, Sunghyun Moon, and Han-Lim Choi. Informative path planning and mapping with multiple UAVs in wind fields. In Distributed Autonomous Robotic Systems, volume 6 of Springer Proceedings in Advanced Robotics, pages 269–283. Springer, 2018. doi: 10.1007/978-3-319-73008-0_19.DOI
  6. [6]Gregory Hitz, Enric Galceran, Marie-Ève Garneau, François Pomerleau, and Roland Siegwart. Adaptive continuous-space informative path planning for online environmental monitoring. Journal of Field Robotics, 34(8):1427–1449, 2017. doi: 10.1002/rob.21722.DOI
  7. [7]Geoffrey A. Hollinger and Gaurav S. Sukhatme. Sampling-based robotic information gathering algorithms. The International Journal of Robotics Research, 33(9):1271–1287, 2014. doi: 10.1177/0278364914533443.DOI
  8. [8]Luzia Knoedler, Oswin So, Ji Yin, Mitchell Black, Zachary Serlin, Panagiotis Tsiotras, Javier Alonso-Mora, and Chuchu Fan. Safety on the fly: Constructing robust safety filters via policy control barrier functions at runtime. IEEE Robotics and Automation Letters, 10(10):10058–10065, 2025. doi: 10.1109/LRA.2025.3597847.DOI
  9. [9]Andreas Krause, Ajit Singh, and Carlos Guestrin. Near-optimal sensor placements in gaussian processes: Theory, efficient algorithms and empirical studies. Journal of Machine Learning Research, 9:235–284, 2008. URL https://jmlr.org/papers/v9/krause08a.html.
  10. [10]Zhongguo Li, Wen-Hua Chen, Jun Yang, and Cunjia Liu. Cooperative active learning-based dual control for exploration and exploitation in autonomous search. IEEE Transactions on Neural Networks and Learning Systems, 36(2):2221–2233, 2025. doi: 10.1109/TNNLS.2024.3349467.DOI
  11. [11]Kai-Chieh Ma, Lantao Liu, and Gaurav S. Sukhatme. Informative planning and online learning with sparse Gaussian processes. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4292–4298, 2017. URL https://arxiv.org/abs/1609.07560.
  12. [12]Roman Marchant and Fabio Ramos. Bayesian optimisation for informative continuous path planning. In 2014 IEEE International Conference on Robotics and Automation (ICRA), 2014. doi: 10.1109/ICRA.2014.6907763.DOI
  13. [13]Brady Moon. Informative Path Planning Toward Autonomous Real-World Applications. Ph.d. thesis, Carnegie Mellon University, Pittsburgh, PA, 2025. URL https://publications.ri.cmu.edu/informative-path-planning-toward-autonomous-real-world-applications.DOI
  14. [14]Kaleb Ben Naveed, Devansh Agrawal, Christopher Vermillion, and Dimitra Panagou. Eclares: Energy-aware clarity-driven ergodic search. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14326–14332, 2024.
  15. [15]Kaleb Ben Naveed, Utkrisht Sahai, Anouck Girard, and Dimitra Panagou. Provably safe stein variational clarity-aware informative planning. In Proceedings of the 8th Annual Learning for Dynamics and Control Conference, volume 331 of Proceedings of Machine Learning Research, pages 1288–1305, 2026. URL https://proceedings.mlr.press/v331/naveed26a.html.
  16. [16]Chaoyi Pan, Zeji Yi, Guanya Shi, and Guannan Qu. Model-based diffusion for trajectory optimization. In Advances in Neural Information Processing Systems, volume 37, 2024. URL https://arxiv.org/abs/2407.01573.
  17. [17]Marija Popović, Teresa Vidal-Calleja, Jen Jen Chung, Juan Nieto, and Roland Siegwart. Informative path planning for active field mapping under localization uncertainty. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10751–10757, 2020a. doi: 10.1109/ICRA40945.2020.9197034.DOI
  18. [18]Marija Popović, Teresa Vidal-Calleja, Gregory Hitz, Jen Jen Chung, Inkyu Sa, Roland Siegwart, and Juan Nieto. An informative path planning framework for UAV-based terrain monitoring. Autonomous Robots, 44:889–911, 2020b. doi: 10.1007/s10514-020-09903-2.DOI
  19. [19]Amarjeet Singh, Andreas Krause, Carlos Guestrin, William J. Kaiser, and Maxim A. Batalin. Efficient planning of informative paths for multiple robots. In Proceedings of the Twentieth International Joint Conference on Artificial Intelligence, pages 2204–2210, 2007. URL https://ijcai.org/proceedings/2007/355.
  20. [20]Dawei Sun, Mohammad Javad Khojasteh, Shubhanshu Shekhar, and Chuchu Fan. Uncertain-aware safe exploratory planning using gaussian process and neural control contraction metric. In Proceedings of the 3rd Conference on Learning for Dynamics and Control, volume 144 of Proceedings of Machine Learning Research, pages 728–741, 2021. URL https://proceedings.mlr.press/v144/sun21a.html.
  21. [21]Mitchell Torok, Man Ching Melvin Chan, Donglin Sui, and Mohammad Deghat. Online wind mapping for coupled path planning and contouring control of quadrotors. Sensors, 26(16):5250, 2026. doi: 10.3390/s26165250.DOI
  22. [22]Andrew Wagenmaker, Guanya Shi, and Kevin G. Jamieson. Optimal exploration for model-based RL in nonlinear systems. In Advances in Neural Information Processing Systems, volume 36, 2023. URL https://papers.nips.cc/paper_files/paper/2023/hash/31e018f43ab9c7065c058cc2c5848128-Abstract-Conference.html.arxiv.org/abs/2306.09210
  23. [23]Andrew J. Wagenmaker, Max Simchowitz, and Kevin Jamieson. Task-optimal exploration in linear dynamical systems. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 10461–10652, 2021. URL https://proceedings.mlr.press/v139/wagenmaker21a.html.arxiv.org/abs/2102.05214

Paper details

vlnc:W_eu9ahh2a
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents