2nd/10_Wiki/Topics/AI_and_ML/Embodied-AI.md

---
id: wiki-2026-0508-embodied-ai
title: Embodied AI
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [embodied AI, robot AI, VLA, vision-language-action, sim2real, robot foundation model]
duplicate_of: none
source_trust_level: A
confidence_score: 0.96
verification_status: applied
tags: [ai, robotics, embodied-ai, vla, foundation-model, sim2real, manipulation]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: ROS 2 / PyTorch / Isaac / MuJoCo
---

# Embodied AI

## 매 한 줄
> **"매 physical body 의 의 perceive + act + learn"**. 매 disembodied LLM 의 X — 매 manipulator + locomotion + navigation. 매 modern: 매 RT-2, OpenVLA, π0 — 매 VLM + action. 매 sim2real + diffusion policy.

## 매 핵심

### 매 task
- **Navigation**: 매 ObjectNav, PointNav.
- **Manipulation**: 매 pick-place, insertion.
- **Locomotion**: 매 quadruped, humanoid.
- **Mobile manipulation**: 매 fetch.
- **Long-horizon**: 매 cook, clean.

### 매 modern method
- **Diffusion Policy** (Chi 2023): 매 visual → action 의 diffusion.
- **VLA** (RT-2, OpenVLA): 매 VLM + action token.
- **π0** (Physical Intelligence): 매 generalist robot foundation.
- **ACT** (Aloha): 매 chunked transformer.

### 매 sim2real
- **Domain randomization**: 매 light, texture, dynamics.
- **Real2sim2real**: 매 real data + sim refine.
- **Co-training**: 매 sim + real mix.

### 매 platform
- **NVIDIA Isaac Sim / Lab**.
- **MuJoCo / DeepMind Control**.
- **PyBullet**.
- **Habitat** (navigation).
- **RoboCasa** (kitchen).

### 매 응용
1. **Industrial**: 매 assembly.
2. **Logistics**: 매 pick-pack.
3. **Service**: 매 cleaning.
4. **Surgery**: 매 da Vinci.
5. **Domestic**: 매 humanoid (1X, Figure, Optimus).

## 💻 패턴

### Diffusion Policy (Chi 2023)
```python
import torch
from torch import nn

class DiffusionPolicy(nn.Module):
    def __init__(self, obs_dim, action_dim, horizon=8, n_steps=100):
        super().__init__()
        self.horizon = horizon
        self.n_steps = n_steps
        self.cond_encoder = nn.Linear(obs_dim, 256)
        self.noise_pred = nn.Sequential(
            nn.Linear(action_dim * horizon + 256 + 1, 512),
            nn.ReLU(),
            nn.Linear(512, action_dim * horizon),
        )
    
    def predict(self, obs):
        cond = self.cond_encoder(obs)
        x = torch.randn(self.horizon * 2)
        for t in reversed(range(self.n_steps)):
            t_emb = torch.tensor([t / self.n_steps])
            noise = self.noise_pred(torch.cat([x, cond, t_emb]))
            x = x - 0.01 * noise
        return x.reshape(self.horizon, -1)
```

### VLA (RT-2 / OpenVLA style)
```python
class VLA(nn.Module):
    def __init__(self, vlm, action_dim=7, n_bins=256):
        super().__init__()
        self.vlm = vlm  # 매 PaLI-X / Llama-VL
        self.action_proj = nn.Linear(vlm.hidden_dim, n_bins * action_dim)
        self.n_bins = n_bins
    
    def forward(self, image, instruction):
        feat = self.vlm(image, instruction).last_hidden_state[:, -1]
        logits = self.action_proj(feat).reshape(-1, 7, self.n_bins)
        action_bins = logits.argmax(-1)
        return self.bin_to_action(action_bins)
```

### Behavior cloning (basic IL)
```python
def behavior_cloning(demos, model):
    """매 (obs, action) 의 supervised learning."""
    optim = torch.optim.AdamW(model.parameters(), lr=1e-4)
    for epoch in range(100):
        for obs, action in demos:
            pred = model(obs)
            loss = F.mse_loss(pred, action)
            optim.zero_grad()
            loss.backward()
            optim.step()
    return model
```

### Sim2Real (domain randomization)
```python
def randomize_env(env):
    env.gravity = np.random.uniform(9.5, 10.1)
    env.friction = np.random.uniform(0.5, 1.5)
    env.light_intensity = np.random.uniform(0.5, 1.5)
    env.texture = random.choice(textures)
    env.payload_mass = np.random.uniform(0, 0.5)
    return env
```

### Habitat navigation
```python
import habitat
config = habitat.get_config('benchmark/nav/objectnav_hm3d_v1.yaml')
env = habitat.Env(config)
obs = env.reset()
while not env.episode_over:
    action = policy(obs)
    obs = env.step(action)
```

### MuJoCo manipulation
```python
import mujoco
model = mujoco.MjModel.from_xml_path('panda.xml')
data = mujoco.MjData(model)
mujoco.mj_step(model, data)
ee_pos = data.site('end_effector').xpos
```

### Reward shaping (manipulation)
```python
def grasp_reward(state):
    distance = np.linalg.norm(state.gripper_pos - state.object_pos)
    in_grasp = state.gripper_holding_object
    lifted = state.object_pos[2] - state.object_init_z > 0.1
    
    return -distance + (5 if in_grasp else 0) + (10 if lifted else 0)
```

### Curriculum learning
```python
def curriculum(success_rate, level):
    if success_rate > 0.8: return level + 1
    if success_rate < 0.3: return max(0, level - 1)
    return level

# 매 level 0: easy (objects close, no obstacles)
# 매 level 1: clutter
# 매 level 2: distractors + dynamic
```

### Real2Sim2Real (RoboCasa-style)
```python
def real2sim(real_traj):
    # 매 real state 의 sim recreate
    sim_init = match_initial_state(real_traj[0])
    sim_traj = simulate(sim_init, real_traj.actions)
    return sim_traj

def sim_train_real_eval(sim_data, real_data):
    model = train_on(sim_data + real_data)
    return evaluate_real(model, real_data.eval)
```

### Action chunking (ACT)
```python
class ACT(nn.Module):
    """매 Aloha bimanual."""
    def __init__(self, chunk=100):
        super().__init__()
        self.chunk = chunk
        self.encoder = TransformerEncoder()
        self.decoder = TransformerDecoder()
    
    def forward(self, obs):
        feat = self.encoder(obs)
        actions = self.decoder(feat)  # 매 [chunk, action_dim]
        return actions
    
    def execute(self, obs):
        chunk = self.forward(obs)
        # 매 temporal ensembling
        return chunk[0]
```

### Safety filter
```python
def safe_action(proposed, state):
    if proposed.force > MAX_FORCE: proposed.force = MAX_FORCE
    if collision_imminent(proposed, state): return STOP_ACTION
    if outside_workspace(proposed, state): return CLAMP_TO_WORKSPACE(proposed)
    return proposed
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Visual policy | Diffusion Policy |
| Language-conditioned | VLA (OpenVLA / π0) |
| Multi-task | Foundation model |
| Long-horizon | Hierarchical + chunking |
| Sim-only | Domain randomization |
| Few demos | BC + augmentation |
| Generalist | π0 / RT-X |

**기본값**: 매 modern = 매 VLA finetune (OpenVLA) + 매 diffusion policy + 매 sim2real domain randomization + 매 safety filter.

## 🔗 Graph
- 부모: [[AI]] · [[Robotics]] · [[Embodied Cognition]]
- 변형: [[VLA]]
- Adjacent: [[Foundation-Model]] · [[CLIP]] · [[π0]]

## 🤖 LLM 활용
**언제**: 매 robot. 매 manipulation. 매 navigation. 매 multimodal physical.
**언제 X**: 매 pure simulation game. 매 disembodied chat.

## ❌ 안티패턴
- **No safety filter**: 매 hardware 의 damage.
- **Sim-only no DR**: 매 sim2real gap.
- **BC overfit demos**: 매 OOD fail.
- **Tiny VLM 의 generalist 의 expect**: 매 capacity 의 부족.
- **No chunking**: 매 jitter / instability.

## 🧪 검증 / 중복
- Verified (RT-2, OpenVLA, Diffusion Policy 2023, π0 2024).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-04-26 | EMBODIED-AI auto |
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — diffusion / VLA / BC / sim2real / curriculum / ACT code |