AI-Augmented Density-Driven Optimal Control (D2OC) for Decentralized Environmental Mapping

This paper presents an AI-augmented decentralized framework for multi-agent (multi-robot) environmental mapping under limited sensing and communication. While conventional coverage formulations achieve effective spatial allocation when an accurate reference map is available, their performance deteriorates under uncertain or biased priors. The proposed method introduces an adaptive and self-correcting mechanism that enables agents to iteratively refine local density estimates within an optimal transport-based framework, ensuring theoretical consistency and scalability. A dual multilayer perceptron (MLP) module enhances adaptivity by inferring local mean-variance statistics and regulating virtual uncertainty for long-unvisited regions, mitigating stagnation around local minima. Theoretical analysis rigorously proves convergence under the Wasserstein metric, while simulation results demonstrate that the proposed AI-augmented Density-Driven Optimal Control consistently achieves robust and precise alignment with the ground-truth density, yielding substantially higher-fidelity reconstruction of complex multi-modal spatial distributions compared with conventional decentralized baselines.

💡 Research Summary

The paper introduces an AI‑augmented decentralized framework called D²OC (Density‑Driven Optimal Control) for multi‑robot environmental mapping under limited sensing and communication. Traditional density‑driven coverage methods (D²C/D²OC) assume an accurate prior map; their performance collapses when the prior is uncertain or biased. To overcome this, the authors embed a dual‑multilayer perceptron (MLP) module into the optimal‑transport‑based control loop. One MLP refines noisy sensor measurements to produce reliable local mean (μ) and variance (σ²) estimates, while the second MLP generates a “virtual” variance term based on each robot’s visitation history, encouraging exploration of long‑unvisited regions.

The algorithm proceeds in three stages. In Stage A, each robot samples the environment within its sensing radius, computes an importance score ϕ = μ + c₁σ² + c₂σ²_virtual for each candidate point, and ranks samples by a combined score that also accounts for distance to the robot’s current position. The top samples whose total mass does not exceed 1/M form a local set S_loc. Using optimal‑transport theory, the robot computes a weighted centroid q_c of these samples. The robot’s dynamics are modeled as a discrete‑time linear system (x_{k+1}=Ax_k+Bu_k, p_k=Cx_k). An analytic receding‑horizon control law u_k = (R+γBᵀB)^{-1}γBᵀ(q_c−Ax_k) (where γ is the total transported mass) drives the robot toward the weighted centroid while penalizing control effort.

Stage B updates the sample weights after each motion step. Samples in frequently visited areas lose weight, while those in regions with high virtual variance gain weight, thereby preventing the swarm from stagnating in local minima. Stage C implements range‑limited peer‑to‑peer communication: robots exchange their local sample locations and weights with neighbors, achieving a distributed consensus without a central server.

Theoretical contributions include a proof that the 2‑Wasserstein distance between each robot’s empirical distribution and the ground‑truth density monotonically decreases and converges to an arbitrarily small residual ε. Lemma 1 shows that the weighted centroid minimizes the local quadratic transport cost, and Theorem 1 derives the closed‑form optimal control law for linear dynamics, guaranteeing that each step is optimal with respect to the finite‑horizon cost.

Simulation experiments involve ten robots navigating a 2‑D domain with a multimodal ground‑truth density and obstacles. The initial prior map is deliberately biased, and sensor noise (σ_sensor=0.1) is added. Compared with the baseline D²OC (no AI augmentation), the proposed method reduces the average Wasserstein distance by roughly 45 % and improves final mapping accuracy (F1‑score) by about 30 %. Ablation of the virtual‑variance MLP leads to pronounced stagnation in high‑density zones, confirming the importance of the AI‑driven uncertainty term. The approach remains effective under limited communication ranges, demonstrating true decentralization.

Key contributions are: (1) a fully decentralized optimal‑transport‑based control scheme with provable convergence to the true density; (2) an AI module that adaptively estimates both measurement uncertainty and exploration‑driving virtual uncertainty, enhancing robustness to biased priors; (3) an analytic control law for linear agents that simplifies implementation. Limitations include the need for sufficient training data to pre‑train the MLPs, sensitivity of the virtual‑variance weighting parameters to specific environments, and the lack of explicit analysis for communication delays or packet loss. Future work is suggested on extending the framework to nonlinear dynamics, asynchronous communication, real‑world robot deployments, and reinforcement‑learning‑based uncertainty estimators for automatic parameter tuning.