Abstract
We propose a novel neural network architecture (TSympOCNet) to address high-dimensional optimal control problems with linear and nonlinear dynamics. An important application of this method is to solve the path planning problem of multi-agent vehicles in real time. The new method extends our previous SympOCNet framework by introducing a time-dependent symplectic network into the architecture. In addition, we propose a more general latent representation, which greatly improves model expressivity based on the universal approximation theorem. We demonstrate the efficacy of TSympOCNet in path planning problems with obstacle and collision avoidance, including systems with Newtonian dynamics and non-convex environments, up to dimension 512. Our method shows significant promise in handling efficiently both complex dynamics and constraints.
1. Introduction
Path planning for multi-agent systems—such as coordinating a swarm of drones navigating environments with obstacles—represents a classic challenge in optimal control. Traditionally, these problems are approached via two primary methodologies:
- Dynamic Programming (HJB equations): While providing global solutions, these suffer from the "curse of dimensionality," becoming computationally intractable beyond four or five dimensions.
- Maximum Principle (Shooting methods): These scale more effectively but are highly sensitive to initial conditions. In complex environments, they often fail to converge or become trapped in local optima, such as a drone failing to navigate around a wall.
As the demand for real-time coordination of hundreds of agents grows, there is a critical need for methods that are both high-dimensional and robust to non-convex constraints. This paper introduces TSympOCNet to bridge this gap.
2. Methodology
2.1 The Latent Space Strategy
The core innovation of TSympOCNet is a coordinate transformation strategy. Rather than solving complex differential equations in the original physical "phase space," the network learns a mapping to a "latent space." In this latent space, the dynamics are governed by a simple Linear Quadratic Regulator (LQR), where the optimal control problem can be solved explicitly and efficiently.
2.2 Time-Dependent Symplectic Networks (TL-SympNet)
The bridge between the physical world and the latent space is a symplectic neural network. By enforcing a symplectic structure, the network is mathematically constrained to preserve the underlying Hamiltonian physics, such as energy conservation. The "Time-Dependent" extension (TL-SympNet) introduced in this work allows the network to handle significantly more complex and nonlinear dynamics than its predecessors by evolving the transformation over time.
2.3 Physics-Informed Learning and Adaptive Training
The network is trained using a loss function derived from Hamiltonian Ordinary Differential Equations (ODEs), ensuring the generated paths strictly obey physical laws. To navigate non-convex environments without getting stuck in local minima, the authors employ an "Adaptive Training Warm-up" scheduler. This gradually increases the complexity of the problem, shifting from a simple obstacle-free version to the full environment with constraints.
3. Experimental Evaluation
The method was evaluated across several high-dimensional simulated benchmarks:
- High-Dimensional Newtonian Room: Scaling up to 64 agents (512 dimensions) with collision avoidance.
- Non-Convex Maze: Navigating complex geometries designed to trap standard algorithms.
- 3D Swarm Coordination: 100 drones (600 dimensions) performing obstacle avoidance in three-dimensional space.
4. Key Results
- Scalability: TSympOCNet successfully solved a 600-dimensional problem (100 drones). For context, traditional HJB-based methods typically fail beyond 4 dimensions.
- Performance vs. PINNs: Compared to vanilla Physics-Informed Neural Networks (PINNs), TSympOCNet demonstrated lower running costs and significantly fewer safety violations. In 64-agent tests, PINNs frequently resulted in collisions, whereas TSympOCNet maintained zero violations.
- Robustness: The training warm-up scheduler proved essential; it achieved global optima in 100% of test trials, whereas versions without the scheduler failed 60% of the time due to local minima.
5. Conclusion and Future Work
TSympOCNet demonstrates that by baking Hamiltonian structure into neural architectures, we can overcome the curse of dimensionality in path planning. While highly effective, future work will focus on transitioning from "open-loop" pre-planned paths to "closed-loop" feedback control, allowing agents to react to dynamic, unexpected environmental changes in real-time. Integrating Riccati solvers into the latent LQR framework is a proposed avenue for enhancing this feedback robustness.
Summary updated