2nd/10_Wiki/Topics/Domain_Programming/AI_and_ML/Physical-Intelligence.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-physical-intelligence	Physical Intelligence	10_Wiki/Topics	verified	self	PI π0 pi-zero embodied-foundation-model	none	A	0.9	applied	robotics foundation-model embodied-ai vla		2026-05-10	pending	language framework Python JAX/PyTorch

Physical Intelligence

매 한 줄

"매 robot 의 universal foundation model — 매 ChatGPT moment for embodied AI". 매 Physical Intelligence (PI, 2024 launch)는 π0 — 매 vision-language-action (VLA) foundation model 의 출시한 startup. 매 single weights 로 매 다양한 robot 매 dishwashing, laundry folding, table bussing 의 수행.

매 핵심

매 회사 + 모델

• 매 회사: Physical Intelligence (Carolina Parada, Sergey Levine, Chelsea Finn 등 — 매 Google Brain/Stanford alumni). 2024 founded, $400M+ raised, $2.4B valuation.
• 매 π0 (pi-zero, 2024-10): 매 first VLA foundation model. PaliGemma (3B VLM) backbone + 매 action expert (300M params, flow matching for continuous actions). 매 50Hz control.
• 매 π0.5 (2025): open-world generalization, hierarchical planning, longer-horizon tasks.
• 매 π0-FAST: tokenized action representation (FAST — Frequency-space Action Sequence Tokenization), 5× faster training.

매 architecture key

• 매 VLA = VLM + action head: 매 vision (ViT) + language (LLM) + action decoder.
• 매 flow matching action expert: 매 continuous robot actions 매 discrete tokens 의 X — 매 flow matching 의 학습.
• 매 cross-embodiment: single model 매 7+ robot platforms (ALOHA, UR5e, Franka, mobile manipulators).
• 매 internet pretraining + robot fine-tune: 매 PaliGemma weights 의 시작 → 매 10K+ hours robot demos 의 training.

매 응용

1. 매 household chore robot (laundry folding, dishwashing).
2. 매 warehouse manipulation (Covariant + PI partnership).
3. 매 humanoid foundation model (Figure 02, 1X NEO compatibility 의 explore).

💻 패턴

π0 inference (lerobot integration)

# pip install lerobot transformers
from lerobot.common.policies.pi0 import PI0Policy
import torch

policy = PI0Policy.from_pretrained("lerobot/pi0")
policy.eval().to("cuda")

obs = {
    "observation.images.top": torch.zeros(1, 3, 224, 224).cuda(),
    "observation.state": torch.zeros(1, 14).cuda(),  # joint positions
    "task": ["fold the towel"],
}

with torch.no_grad():
    action_chunk = policy.select_action(obs)  # (1, 50, 14) — 50-step chunk

Action chunking + temporal ensembling

# π0 outputs 50-step action chunks at 50Hz
# Execute first k steps, predict again — temporal ensemble for smoothness
ACTION_HORIZON = 50
EXECUTE_STEPS = 8

action_buffer = collections.deque(maxlen=ACTION_HORIZON)
for t in range(MAX_STEPS):
    if t % EXECUTE_STEPS == 0:
        chunk = policy(obs)  # predict 50-step chunk
        action_buffer.extend(chunk[0])
    action = action_buffer.popleft()
    obs = robot.step(action)

Flow matching action head (simplified)

import torch.nn as nn

class FlowMatchingActionHead(nn.Module):
    def __init__(self, dim=1024, action_dim=14, horizon=50):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(dim + action_dim + 1, 1024), nn.SiLU(),
            nn.Linear(1024, action_dim),
        )

    def forward(self, vlm_features, noisy_action, t):
        x = torch.cat([vlm_features, noisy_action, t], dim=-1)
        return self.net(x)  # velocity field

    def sample(self, vlm_features, num_steps=10):
        a = torch.randn(B, 50, 14)
        for i in range(num_steps):
            t = i / num_steps
            v = self.forward(vlm_features, a, t)
            a = a + v / num_steps
        return a

LeRobot dataset format

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("lerobot/aloha_static_fork_pick_up")
sample = ds[0]
# {'observation.images.top': tensor, 'observation.state': tensor,
#  'action': tensor, 'task': 'pick up the fork'}

Cross-embodiment fine-tune

# Fine-tune π0 on a new robot (e.g. custom 6-DoF arm)
config = PI0Config(action_dim=6, state_dim=6)  # adjust dims
policy = PI0Policy(config)
policy.load_pretrained_vlm("lerobot/pi0")  # load PaliGemma + freeze
# Train action head only on small (~1000 episode) custom dataset

Language-conditioned task switch

# Same weights, different language prompts → different behaviors
for task in ["fold the shirt", "pick up the cup", "wipe the table"]:
    obs["task"] = [task]
    action = policy(obs)
    execute(action)

매 결정 기준

상황	Approach
매 single-task robot, abundant data	매 task-specific BC/RL
매 multi-task, language-conditioned	매 π0 fine-tune
매 zero-shot new task	매 π0.5 (open-world)
매 humanoid full-body	매 π0 + whole-body controller
매 high-frequency control (>100Hz)	매 distill π0 → smaller policy

기본값: 매 cross-embodiment manipulation 의 π0 fine-tune (lerobot 사용).

🔗 Graph

• 부모: Embodied-AI · Foundation-Models

🤖 LLM 활용

언제: 매 multi-task robot manipulation, language-conditioned policy, cross-embodiment transfer 의 사용. 언제 X: 매 simple pick-and-place (overkill), 매 sub-50Hz needed (latency), 매 contact-rich precision tasks (still ongoing research).

❌ 안티패턴

• 매 raw pretrained π0 deploy: 매 fine-tune 없이 — 매 robot/scene mismatch 의 fail.
• 매 ignore action chunking: 매 single-step prediction → 매 jittery motion.
• 매 mismatched camera intrinsics: 매 training cam 매 deploy cam 의 different → 매 OOD failure.

🧪 검증 / 중복

• Verified (Physical Intelligence official, π0 paper 2024-10, lerobot integration).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — π0/π0.5 VLA foundation model + lerobot patterns