A Dual Curriculum Learning Framework for Multi-UAV Pursuit-Evasion in Diverse Environments

TL;DR

The paper tackles cooperative pursuit-evasion by multiple UAVs in diverse 3D environments with obstacles. It introduces DualCL, a dual curriculum learning framework combining an Intrinsic Parameter Curriculum Proposer and an External Environment Generator, built on MAPPO, to address vast exploration and drone-dynamics challenges. Empirical results show DualCL achieving over 90% capture rates, reducing capture time by at least 27.5%, strong zero-shot generalization to unseen scenarios, and viable sim-to-real transfer demonstrated on real quadrotors. This work advances practical, cooperative UAV pursuit strategies in complex, realistic environments with demonstrable generalization and deployment potential.

Abstract

This paper addresses multi-UAV pursuit-evasion, where a group of drones cooperates to capture a fast evader in a confined environment with obstacles. Existing heuristic algorithms, which simplify the pursuit-evasion problem, often lack expressive coordination strategies and struggle to capture the evader in extreme scenarios, such as when the evader moves at high speeds. In contrast, reinforcement learning (RL) has been applied to this problem and has the potential to obtain highly cooperative capture strategies. However, RL-based methods face challenges in training for complex 3-dimensional scenarios with diverse task settings due to the vast exploration space. The dynamics constraints of drones further restrict the ability of reinforcement learning to acquire high-performance capture strategies. In this work, we introduce a dual curriculum learning framework, named DualCL, which addresses multi-UAV pursuit-evasion in diverse environments and demonstrates zero-shot transfer ability to unseen scenarios. DualCL comprises two main components: the Intrinsic Parameter Curriculum Proposer, which progressively suggests intrinsic parameters from easy to hard to improve the capture capability of drones, and the External Environment Generator, tasked with exploring unresolved scenarios and generating appropriate training distributions of external environment parameters. The simulation experimental results show that DualCL significantly outperforms baseline methods, achieving over 90% capture rate and reducing the capture timestep by at least 27.5% in the training scenarios. Additionally, it exhibits the best zero-shot generalization ability in unseen environments. Moreover, we demonstrate the transferability of our pursuit strategy from simulation to real-world environments. Further details can be found on the project website at https://sites.google.com/view/dualcl.

Figures (6)

• Figure 1: Workflow of our dual curriculum learning framework. Taking the phase $i$ as an example, we utilize the Intrinsic Parameter Curriculum Proposer to generate the intrinsic parameters $\tau^i_{int}$ for the current phase and the External Environment Generator to produce external environment parameters $\tau^i_{ext}$. We combine the intrinsic and external environment parameters to form the task parameters set $\mathcal{M}^i$ and generate parallel simulation environments for MARL training.
• Figure 2: The aerial view of the multi-UAV pursuit-evasion problem. In (a) and (b), we demonstrate two infeasible scenarios and two feasible scenarios. In (c), we show the grid pattern of the scenario and verify its solvability using DFS.
• Figure 3: Results of DualCL and all baselines in Empty. DualCL achieves $100\%$ capture rate and the shortest capture timestep with all combinations of intrinsic parameters.
• Figure 4: Visualization of the capture strategies generated by DualCL and APF.
• Figure 5: Ablations on two main components.