This directory contains the configurations and fixed policies for the 130 generalization evaluations: ten policies evaluated on thirteen conditions. The same policy is used across conditions without retraining.
| Contents | Purpose |
|---|---|
| configs/ | The nominal environment and twelve perturbations: target positions (a1–a3), rear-row obstacle positions (b1–b3), retained obstacles (c1–c3), and flow speeds (d1–d3) |
policies/SEAS_1/ through SEAS_5/ |
Five executable Julia controllers selected from the evolution runs |
policies/DRL_1/ through DRL_5/ |
Five PPO model checkpoints with their observation-normalization and random-state files |
| policies/index.json | Controller origins, checkpoint selection, file identities and the recorded inference mode for each policy |
For a DRL checkpoint, model.zip contains the model, vecnormalize.pkl contains
the observation-normalization statistics, and rng_state.pkl records the
checkpoint's random state. DRL evaluations use deterministic mean
actions (deterministic=True) through agent_fluid_ppo.evaluate_frozen_policies.
The shared implementations are the 2D CFD simulator and the DRL training and inference package. Recorded evaluation outcomes and trajectories are in raw_data/generalization/. The Figure 3 replay records contain the selected flow fields for four illustrated cases.
From the repository root, the extraction script reconstructs these results together with the other manuscript data. The processed Table 1 inputs and Figure 3 inputs are used by the table generator and Figure 3 plotting script.