hwihwalab/neuromotion-humanoid-v5-ppo

🤖 NeuroMotion // Humanoid-v5 PPO Continuous Control

MuJoCo Humanoid-v5 Continuous Control Telemetry & PPO Reinforcement Learning System
🌐 English Documentation | 🇰🇷 한국어 매뉴얼

This repository contains an advanced continuous reinforcement learning system (PPO) and a real-time engineering telemetry dashboard for bipedal robotic continuous control in Gymnasium MuJoCo Humanoid-v5.

---

📊 Model Specifications & Benchmark Results

Parameter	Specification
Environment	Gymnasium MuJoCo Humanoid-v5
Observation Space	378-dimensional continuous vector
Action Space	17-dimensional continuous joint torques (Box[-1.0, 1.0])
Algorithm	Proximal Policy Optimization (PPO)
Framework	Stable-Baselines3 / PyTorch
Architecture	Actor-Critic MLP Policy (MlpPolicy)
Average Survival Steps	88.5 steps (Peak: 109 steps)
Average Cumulative Reward	+455.52 (Peak: +549.18)

---

🚀 Quick Start (Inference & Evaluation)

1. Download & Load with Stable-Baselines3

import gymnasium as gym
from stable_baselines3 import PPO
from huggingface_hub import hf_hub_download

# Download model weights from Hugging Face Hub
model_file = hf_hub_download(
    repo_id="hwihwalab/neuromotion-humanoid-v5-ppo",
    filename="humanoid_ppo_model.zip"
)

# Initialize MuJoCo Humanoid-v5 environment
env = gym.make("Humanoid-v5", render_mode="human")
model = PPO.load(model_file, env=env)

# Run evaluation episodes
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()

---

🖥️ Real-time Telemetry Dashboard (NeuroMotion Studio)

The repository also includes a native 60fps Pygame telemetry cockpit:

python run_gui.py

Features:

• 7-Panel Telemetry Layout: 17-DOF Actuator Torque Bars, Learning Diagnostics (Policy/Value Loss, Entropy), Reward Telemetry Curves, Top Bento KPI Cards.
• Interactive Controls: Real-time Physics Speed Multipliers (1x, 2x, 4x, 8x), Disturbance Force Injection (15N, 30N, 50N), Checkpoint Save/Load.

---

📂 Repository Contents

• README.md: Hugging Face model card, specifications, and telemetry manual.
• README_KR.md: Dedicated Korean manual (한국어 매뉴얼).
• humanoid_ppo_model.zip: Pre-trained PPO neural network weights.
• simulation_engine.py: Gymnasium MuJoCo physical engine & 17-DOF torque extractor.
• train_rl.py: Stable-Baselines3 PPO incremental trainer & diagnostics.
• run_gui.py: Pygame real-time telemetry cockpit.
• mujoko_humanoid.py: Multi-mode CLI execution and visualization script.

---

Trained and deployed with NeuroMotion Studio by hwihwalab.