Problem

The primary challenge addressed by this research is the need for a highly effective and reliable Automatic Ground Collision Avoidance System (AGCAS) for advanced jet trainers. Existing systems require enhancement to improve operational safety and efficiency, particularly in situations where pilot intervention is not feasible or timely. Developing an AI-driven system capable of robustly avoiding terrain collisions, maintaining stable flight, and minimizing control oscillations within a limited observation space is crucial for modern aerospace operations.

Method

The researchers developed an AGCAS that employs a reinforcement learning (RL) model, specifically a customized Soft Actor-Critic (SAC) algorithm. A custom Convolutional Neural Network (CNN) was integrated into the SAC algorithm's feature extractor layer to process visual and sequential state representations efficiently. This CNN utilizes "pseudo-lidar" data, which combines Height-on-Terrain (HoT) and Height-above-Terrain (HaT) from a digital elevation map (DEM) terrain server, along with RADALT inputs and aircraft base states. A sequential reward function was designed to prioritize collision avoidance (with a significant -250 penalty for collisions), maintain wing-level flight, and minimize oscillating control actions. Various LiDAR input configurations (single point, 3x5, 5x5, and KxK arrays like 16x16) were experimented with, and hyperparameter optimization was performed using Optuna to achieve optimal performance.

Results

The final CNN-SAC model, particularly with a 16x16 LiDAR input, demonstrated superior performance compared to other tested configurations. The integration of the CNN into the SAC feature extraction layer significantly enhanced the agent's ability to interpret complex terrain data, identify potential escape routes, and plan valid collision avoidance paths. The system proved to be more generalizable, stable, and reliable, improving upon traditional control-based AGCAS. Expert validation and simulation data confirmed the effectiveness of the AI-driven approach in significantly improving AGCAS performance for advanced jet trainers.

Implications

This research offers significant implications for enhancing safety and operational capabilities in advanced jet trainers through the application of sophisticated AI. The successful integration of reinforcement learning with a custom CNN for processing complex visual and sequential data provides a robust framework for future advancements in autonomous aviation systems. The methodology can be extended to other critical control applications where real-time obstacle avoidance and stable control are paramount, potentially reducing pilot workload and increasing mission success rates in high-stress environments.