CptAdolito69/ppo-seals-Hopper-v1

PPO Agent playing seals/Hopper-v1

This is a trained model of a PPO agent playing seals/Hopper-v1 using the stable-baselines3 library.

The agent was trained for ~18M timesteps and reaches a mean episodic reward of 3333.21 ± 354.25 over the last 100 training episodes (episodes have a fixed horizon of 1000 steps, as defined by the seals suite).

Demo

Usage (with Stable-Baselines3)

from huggingface_sb3 import load_from_hub
from stable_baselines3 import PPO

# Download the model from the Hub
checkpoint = load_from_hub(
    repo_id="CptAdolito69/ppo-seals-Hopper-v1",
    filename="ppo-seals-Hopper-v1.zip",
)
model = PPO.load(checkpoint)

# Run it (requires `seals` + a MuJoCo backend installed)
import gymnasium as gym
import seals  # noqa: F401  registers the seals/* environments

env = gym.make("seals/Hopper-v1")
obs, _ = env.reset(seed=42)
done = False
while not done:
    action, _ = model.predict(obs, deterministic=True)
    obs, reward, terminated, truncated, _ = env.step(action)
    done = terminated or truncated

Training details

Hyperparameter	Value
Algorithm	PPO
Policy	MlpPolicy
Environment	seals/Hopper-v1
Timesteps	17,989,632
Learning rate	3e-4
Seed	42
Observation space	Box(12,)
Action space	Box(3,)
SB3 version	2.6.0
PyTorch version	2.7.1
Gymnasium version	1.1.1

Files in this repo

• ppo-seals-Hopper-v1.zip — the loadable SB3 model (PPO.load(...))
• policy.pth, policy.optimizer.pth, pytorch_variables.pth, data — extracted model contents
• config.json — training configuration summary
• replay.mp4 — demo rollout of the trained agent
• system_info.txt — environment used for training