PPO Multi-Seed Instability Dataset and Analysis Code

This dataset and accompanying analysis code support the empirical study titled "Reproducibility Failures in Proximal Policy Optimization: Multi-Seed Instability, Policy Collapse, and Reward–Behaviour Mismatch in 3D Drone Navigation." The dataset comprises two CSV files generated from training a Proximal Policy Optimization (PPO) agent across five independent random seeds (0–4) for 1,000 episodes each (5,000 total episodes, 1,500,000 environment steps) in a custom 3D drone point-navigation environment. The file ppo_train.csv contains episode-level training logs including seed identifier, episode number, total reward, mean policy entropy, and task success flag across all 5,000 episodes. The file ppo_trajectory.csv contains step-level spatial logs recording the agent's (x, y, z) position and per-step reward at each of the 300 steps per episode, totalling 1,500,000 rows. The environment task requires navigating from a fixed start position (0, 0, 0.5) to a goal at (10, 10, 5.0) within a bounded 3D space using a progress-based shaped reward. The repository also includes the Python training script (ppo_final.py), three R analysis scripts (Final_ppo.R, entropy_diagnostic.R, baseline_comparison.R), and a simulated random policy dataset (random_policy_results.csv) used as a baseline comparison. All code is fully reproducible using the provided random seeds.