ftle-informed planning

This work addresses the challenge of multi-agent rendezvous in complex fluid environments, where agents often become trapped in separated states due to vortex dynamics. The authors propose a fluid physics-informed multi-agent reinforcement learning (MARL) strategy that breaks state-action mapping symmetry to uncover non-intuitive cooperative mechanisms, effectively avoiding vortex-induced traps. The approach demonstrates strong generalization across varying vortex intensities, spatial scales, and group sizes, and yields heuristic policies that outperform baseline methods. Experimental results show that the proposed MARL strategy significantly improves rendezvous success rates. Furthermore, theoretical analysis reveals that fluid deformation impedes the rendezvous process, suggesting that regions with weaker deformation are more suitable as rendezvous targets.

This work addresses the limitations of existing flow matching methods, which rely on straight-line trajectories and struggle to capture complex dynamical behaviors. By leveraging the principle of least action, the authors generalize the Lagrangian formalism to construct probability paths and associated velocity fields that satisfy both the continuity equation and endpoint constraints. This approach embeds flow matching within a classical mechanics framework, thereby unifying and extending prior methods such as optimal transport and diffusion-based paths. The resulting static variational objective eliminates the need for trajectory simulation and enables direct optimization. Empirical results demonstrate that the proposed method yields physically meaningful dynamical evolutions and achieves performance on par with state-of-the-art conditional flow matching models.

This work addresses the challenge of autonomous navigation in unsteady, time-varying fluid environments characterized by partial observability and inherent unpredictability. The authors propose a navigation strategy based on Twin Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning that relies solely on local flow velocity, vorticity, and short-term memory to achieve goal-directed navigation within a parameterized chaotic double-gyre flow field. Systematic evaluation of various bio-inspired sensing modalities reveals that explicit access to global flow parameters degrades performance, whereas agents integrating local velocity and vorticity perception achieve optimal results: velocity sensing enhances energy efficiency, while vorticity sensing improves mapping of flow structures and accuracy in target approach. These findings demonstrate that robust navigation in complex flows is better supported by implicit, local sensory cues rather than global flow information.

This work addresses the lack of theoretical guarantees for autonomous exploration in time-varying flow fields by proposing an ergodic coverage method that explicitly incorporates environmental dynamics. It extends the Maximum Mean Discrepancy (MMD) ergodic metric to evolving domains, formulating a coverage objective function that embeds flow field dynamics. The approach achieves asymptotically optimal coverage under open-loop and underactuated constraints. By integrating MMD-based ergodic control, flow-adaptive path planning, and non-convex dynamic environment optimization, the method demonstrates effectiveness in applications such as oceanic surveying and tracking of human or animal movements. Real-world experiments with aerial and legged robots successfully validate its capability to achieve efficient ergodic coverage in non-convex, dynamic flow environments.

This work addresses real-time trajectory optimization and cooperative control of autonomous agents on resource-constrained edge devices. We propose an efficient Model Predictive Control (MPC) framework based on integral Chebyshev collocation. Our key contribution is the first integration of integral Chebyshev polynomial parameterization with differentiable polyhedral collision checking, enabling explicit modeling of actuator saturation and hard obstacle-avoidance constraints. The formulation minimizes L₂ approximation error and is solved via quadratic programming for rapid convergence. Employing a receding-horizon MPC architecture, the method achieves over 3.2× speedup over conventional approaches on edge hardware, enabling sub-millisecond replanning. We validate its safety, real-time performance, and cooperative capability in multi-spacecraft formation control tasks, demonstrating robust constraint satisfaction and scalable coordination under tight computational budgets.

This study addresses the challenge of detecting transient chaos and abrupt dynamical transitions in equation-free scalar time series by proposing a geometry-guided method for estimating finite-time Lyapunov exponents (FTLE). The approach uniquely integrates a structural proximity matrix derived from Poincaré section occupancy grids with predicted trajectory divergence, introduces macroscopic spatial discretization as a topological regularizer, and employs k-nearest-neighbor extrapolation errors to estimate local instability. Geometric latent variables aligned with empirical FTLE spectra are extracted via partial least squares regression. Experimental results demonstrate that, without access to governing equations, the method accurately tracks asymptotically damped dynamics and sudden phase-space collapses, significantly outperforming the QR-FTLE baseline under moderate signal-to-noise ratios while enhancing detection accuracy for continuous transitions and robustness to noise.

This work addresses the challenge of collision-free trajectory planning for large-scale omnidirectional floating robot formations on water, where strong coupling and combinatorial complexity hinder scalability. The authors propose an extensible planning pipeline that decomposes the coupled problem into interaction clusters via a collision graph, enabling parallel trajectory optimization within each cluster. A robust mechanism handles decomposition failures, and the framework supports real-time interactive editing under sparse keyframe constraints. This approach achieves, for the first time, the generation of smooth, minute-long, collision-free trajectories for hundreds of robots within seconds. The method is validated in simulations with up to 500 agents and successfully deployed in real-world scenarios, including a 24-boat formation on Lake Zurich and the 2025 Venice Architecture Biennale, thereby overcoming the scalability bottleneck in large-scale aquatic coordinated motion planning.

This work addresses the challenge of achieving real-time, multimodal obstacle avoidance with kinematic consistency in dense, highly dynamic environments—a task where existing methods often fall short. The authors propose a physics-informed rectified flow–based policy distillation framework that encodes a model predictive control (MPC) expert policy into a continuous-time ordinary differential equation (ODE). By integrating parallel latent sampling, lightweight feasibility filtering, and an asynchronous action-chunking architecture, the method achieves millisecond-level single-step inference while preserving kinematic consistency. Experiments demonstrate a 98.85% success rate with zero collisions in simulation, at an average inference latency of just 1.29 ms—yielding a 37.2× speedup over MPC and an 800× improvement over standard diffusion models. On real-world edge hardware, the system maintains stable operation with approximately 5.3 ms latency.

This work addresses coordination in structured multi-agent transportation systems where agents follow fixed, non-reroutable paths and must schedule their passage through waypoints to avoid collisions while maximizing efficiency. The authors propose a nonlinear speed scheduling method that eliminates the need for integer ordering variables by employing a differentiable trajectory model to map time into smooth position curves. Safety constraints based on inter-agent distances are enforced over a dense temporal grid. An inexact projected ADMM algorithm efficiently solves the resulting optimization problem by integrating structured temporal updates with gradient-based collision-avoidance corrections. Experiments demonstrate that the approach consistently generates feasible and efficient schedules across diverse scenarios—including random intersections, bottlenecks, and graph-structured networks—with notably improved makespan performance in bottleneck settings compared to representative hierarchical baselines.

本文针对非线性系统中的最优控制问题，提出了一种基于分散前向树搜索的规划算法DFT*，并证明了其在有限样本下的近似最优性。