safe-autonomous-systems/ppo-RBC2D-hard-v0

PPO on RBC2D-hard-v0 (FluidGym)

This repository is part of the FluidGym benchmark results. It contains trained Stable Baselines3 agents for the specialized RBC2D-hard-v0 environment.

Evaluation Results

Global Performance (Aggregated across 5 seeds)

Mean Reward: -0.31 ± 0.03

Per-Seed Statistics

Run	Mean Reward	Std Dev
Seed 0	-0.32	2.24
Seed 1	-0.27	2.19
Seed 2	-0.36	2.50
Seed 3	-0.29	2.22
Seed 4	-0.30	2.35

About FluidGym

FluidGym is a benchmark for reinforcement learning in active flow control.

Usage

Each seed is contained in its own subdirectory. You can load a model using:

from stable_baselines3 import PPO
model = PPO.load("0/ckpt_latest.zip")

**Important:** The models were trained using ```fluidgym==0.0.2```. In order to use
them with newer versions of FluidGym, you need to wrap the environment with a
`FlattenObservation` wrapper as shown below:
```python
import fluidgym
from fluidgym.wrappers import FlattenObservation
from stable_baselines3 import PPO

env = fluidgym.make("RBC2D-hard-v0")
env = FlattenObservation(env)
model = PPO.load("path_to_model/ckpt_latest.zip")

obs, info = env.reset(seed=42)

action, _ = model.predict(obs, deterministic=True)
obs, reward, terminated, truncated, info = env.step(action)

References

• Plug-and-Play Benchmarking of Reinforcement Learning Algorithms for Large-Scale Flow Control
• FluidGym GitHub Repository