Reinforcement Learning
stable-baselines3
English
Korean
ppo
gymnasium
mujoco
pusher-v5
robotics
Eval Results (legacy)
Instructions to use leegoheun/ppo-pusher-v5 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- stable-baselines3
How to use leegoheun/ppo-pusher-v5 with stable-baselines3:
from huggingface_sb3 import load_from_hub checkpoint = load_from_hub( repo_id="leegoheun/ppo-pusher-v5", filename="{MODEL FILENAME}.zip", ) - Notebooks
- Google Colab
- Kaggle
PPO Agent Trained on Gymnasium MuJoCo Pusher-v5
μ΄ λͺ¨λΈμ Gymnasium MuJoCo Pusher-v5 νκ²½μμ νμ΅λ Stable-Baselines3 PPO (Proximal Policy Optimization) κ°ννμ΅ λ‘λ΄ ν μ μ΄ μ μ±
λͺ¨λΈμ
λλ€.
π€ νκ²½ μ 보 (Environment Specs)
- νκ²½:
Pusher-v5(MuJoCo 7-DoF Robotic Arm) - λͺ©ν: 7κ° κ΄μ λͺ¨ν°λ₯Ό μ μ΄νμ¬ ν μ΄λΈ μμ μν΅ λ¬Όμ²΄λ₯Ό λͺ©ν μ§μ (Goal)μΌλ‘ λ°μ΄ λ£κΈ°
- μν κ³΅κ° (23-dim): κ΄μ κ°λ(7), κ΄μ μλ(7), λ‘λ΄ μλ(3), 물체(3), λͺ©ν μμΉ(3)
- νλ κ³΅κ° (7-dim): κ° κ΄μ λͺ¨ν°μ μΈκ°λλ μ°μ ν ν¬ $[-1.0, 1.0]$
- μ΅κ³ λ¬μ± 보μ:
-39.05(μ΄100,000νμμ€ν νμ΅)
π λͺ¨λΈ λ‘λ λ° μΆλ‘ μ¬μ©λ² (Usage)
import gymnasium as gym
from stable_baselines3 import PPO
from huggingface_hub import hf_hub_download
# 1. νκΉ
νμ΄μ€μμ λͺ¨λΈ λ€μ΄λ‘λ
model_path = hf_hub_download(repo_id="leegoheun/ppo-pusher-v5", filename="ppo_pusher.zip")
# 2. νκ²½ μμ± λ° λͺ¨λΈ λ‘λ
env = gym.make("Pusher-v5", render_mode="human")
model = PPO.load(model_path, env=env)
# 3. λ‘λ΄ ν μ μ΄ μ€ν
obs, info = env.reset()
for _ in range(1000):
action, _states = model.predict(obs, deterministic=True)
obs, reward, terminated, truncated, info = env.step(action)
if terminated or truncated:
obs, info = env.reset()
env.close()
π νμ΅ νλΌλ―Έν° (Hyperparameters)
Algorithm: PPO
Policy: MlpPolicy (23 -> 64 -> 64 -> 7)
Learning Rate: 3e-4
Timesteps: 100000
Batch Size: 64
n_steps: 2048
Gamma: 0.99
GAE Lambda: 0.95
Clip Range: 0.2
- Downloads last month
- 25
Evaluation results
- Mean Evaluation Reward on Pusher-v5self-reported-39.050