Dynamic-Target Potential Pursuit Field Reward for UAV Reinforcement Learning

The proposed PPF-based hierarchical reinforcement learning framework

This is the official for manuscript entitled Dynamic-Target Pursuit Potential Field Reward for UAV Reinforcement Learning submitted to IEEE Transactions on Control Systems Technology.

Potential Pursuit Field (PPF), a novel reward shaping framework aimed to address the reward sparsity in reinforcement learning for dynamic target pursuit. By designing a droplet-shaped anisotropic potential field, the proposed PPF model provides dense and direction-aware reward signals while preserving policy invariance through potential-based reward shaping. Building upon PPF, we developed a hierarchical reinforcement learning algorithm, enabling target pursuit and obstacle avoidance in non-line-of-sight(NLOS) environments, simultaneously.

Video Demo

🎬Bilibili

🎬 Watch demo (MP4)

Potential Pursuit Field(PPF)

Art work of Potential Pursuit Field (PPF)

A novel concept of the Potential Pursuit Field (PPF) is proposed to support a continuous and dense reward-shaping function, which can capture anisotropic features and obtain richer gradient information than that of traditional rewards.

Obstacle-free pursuit

Pursuit with Obstacle envirionment

The entire code and corresponding simulation environment will be released later.