Abstract
High-speed drone path planning requires balancing safety and energy efficiency under strict real-time constraints. Traditional methods like Reinforcement Learning often fail to converge quickly enough for emergency obstacle avoidance. This paper introduces RUDE (RUles and DEep learning), a hybrid framework that combines geometric rules with a deep learning-based Situation Prediction Model (SPM). By using rules to constrain the search space and deep learning to optimize trajectory selection, RUDE achieves a 99.6% success rate in navigating restricted zones while preserving critical kinetic energy.
Introduction
Path planning for drones traveling at several kilometers per second is a high-stakes optimization problem. When a "no-fly zone" suddenly appears, the drone must veer off-course to avoid it without depleting the kinetic energy required to reach its final destination. This balancing act is NP-hard. Existing algorithms, such as Evolutionary Algorithms (EAs), often struggle with the massive search space or fail to meet real-time decision-making requirements.
The RUDE Framework
The RUDE framework addresses these challenges by fusing expert-designed rules with neural networks. The process is divided into three main stages:
- State Reduction and Candidate Generation: Instead of continuous steering, RUDE simplifies the path into a set of "middle way-points." An expert rule-based system calculates a safe baseline solution, and 1,000 candidate way-point sets are generated randomly around this baseline.
- Situation Prediction Model (SPM): The core of the policy is the SPM, which consists of two Multi-Layer Perceptron (MLP) networks. The Discriminator predicts the probability of zone intrusion, while the Energy Predictor estimates kinetic energy consumption for each candidate path.
- Policy Selection: During flight, the drone evaluates all 1,000 candidates through the SPM and selects the trajectory that minimizes energy loss while maintaining near-zero intrusion probability.
Experimental Setup
The researchers utilized a custom simulation platform with a realistic physics engine. The training dataset was generated by varying starting positions (600–1300 km), initial speeds (4–6 km/s), and restricted zone radii (100–250 km). The SPM was trained using cross-entropy loss for the discriminator and Mean Squared Error (MSE) for the energy predictor. To improve generalization, the authors implemented a "symmetry trick," establishing a coordinate system relative to the restricted zone's center, which simplifies the state space regardless of global coordinates.
Results and Discussion
RUDE demonstrates a significant performance leap over traditional rule-based methods. In comparative testing, RUDE achieved a success rate of 99.6%, whereas the expert-designed baselines, Rule1 and Rule2, only reached 41.8% and 47.7%, respectively.
The model's success is attributed to its ability to use rules as "safety rails." By constraining the AI to evaluate only "reasonable" paths near a rule-based baseline, the system avoids the aimless exploration typical of pure deep learning approaches. Furthermore, RUDE decouples data collection from training, allowing for more efficient, parallelized model development compared to standard Reinforcement Learning.
Conclusion
The RUDE framework provides a robust solution for high-speed drone navigation in the presence of restricted zones. While currently focused on single-obstacle scenarios, the approach paves the way for multi-obstacle avoidance and more complex vehicle dynamics. Future work will involve testing the framework against dynamic obstacles and validating performance on physical hardware in high-speed environments.
Summary updated