Instructions to use DAVIAN-Robotics/pi05-robocasa-H50 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LeRobot
How to use DAVIAN-Robotics/pi05-robocasa-H50 with LeRobot:
- Notebooks
- Google Colab
- Kaggle
pi0.5 โ RoboCasa H50 (fine-tuned)
๊ฐ์ sk-260713-vision-lecture (SK ๋น์ ๊ฐ์, ์๋ฎฌ๋ ์ด์
VLA ์ค์ต) ์ baseline ์ฒดํฌํฌ์ธํธ๋ค.
LeRobot ์ pi05 ์ ์ฑ
(ฯ0.5, flow-matching VLA)์
lerobot/pi05_base ์์ ์ถ๋ฐํด RoboCasa ํค์น ๋ฐ์ดํฐ์
์ผ๋ก full fine-tuning ํ๋ค.
๋ฐ์ดํฐ
DAVIAN-Robotics/robocasa-H50โ 1,293 ์ํผ์๋ / 242 ํ์คํฌ (LeRobot v3.0 ํฌ๋งท, RoboCasa ํค์น).
ํ์ต ์ค์
| ํญ๋ชฉ | ๊ฐ |
|---|---|
| base | lerobot/pi05_base (PaliGemma gemma_2b + action expert gemma_300m) |
| learning rate | 1e-4 (cosine decay, warmup 3,000 โ decay_lr 2.5e-6) |
| batch size | 64 |
| steps | 60,000 |
| tune targets | tune_all (visual + llm + action_expert ์ ๋ถ ํ์ต) |
| chunk size | 50 |
| dtype | bfloat16 |
| optimizer | AdamW (betas 0.9/0.95, wd 0.01, grad clip 1.0) |
๋ชจ๋ธ config (config.json)
chunk_size = 50,n_action_steps = 25,num_inference_steps = 10(flow matching ์ ๋ถ ์คํ )- ์นด๋ฉ๋ผ 3๋ โ
robot0_agentview_left_image,robot0_agentview_right_image,robot0_eye_in_hand_image. ๊ฐ3ร128ร128์ ๋ ฅ์ด๋ฉฐ ์ ์ฑ ๋ด๋ถ์์224ร224(image_resolution) ๋ก ๋ฆฌ์ฌ์ด์ฆ๋๋ค. - state 16-d (
max_state_dim=32๋ก ํจ๋ฉ), action 12-d (max_action_dim=32๋ก ํจ๋ฉ) - ์ ๊ทํ: VISUAL=IDENTITY, STATE/ACTION=QUANTILES โ ํต๊ณ๋
policy_preprocessor_*.safetensors/policy_postprocessor_*.safetensors์ baked ๋์ด ์๋ค. ๋ฐ๋์ ํจ๊ป ๋ก๋ํด์ผ ํ๋ค (์๋ ์ฐธ์กฐ).
๊ฒ์ฆ๋ ์ฑ๋ฅ
PnPCounterToCab, 40 ์ํผ์๋, max_steps=500:
| ์กฐ๊ฑด | Success Rate |
|---|---|
| ์๋ณธ ๋๋ค ์ฌ (layout/style/์ค๋ธ์ ํธ ์ ๋ถ ๋๋ค โ ํ์ต ๋ถํฌ์ ๋์ผ) | 0.550 |
| ๊ณ ์ -apple ์ฌ (layout 0 / style 0 / apple โ ๊ฐ์ ์ค์ต ๊ธฐ๋ณธ config) | 0.575 |
ฯ0.5 ๋ flow matching ์ด๋ผ ์ฒญํฌ๋ง๋ค ๋ ธ์ด์ฆ๋ฅผ ์๋ก ๋ฝ๋๋ค โ ๊ฐ์ env seed ์์๋ SR ์ด ํ๋ค๋ฆฐ๋ค. 40 ์ํผ์๋ ๊ธฐ์ค 95% Wilson CI ๋ ์ฝ ยฑ0.15 ๋ค. ์ด ๋ฒ์ ๋ฐ์ ์ฐจ์ด๋ง ์ ์๋ฏธํ๊ฒ ์ฝ์ด๋ผ.
์ฌ์ฉ๋ฒ
from lerobot.policies.factory import get_policy_class, make_pre_post_processors
from huggingface_hub import snapshot_download
path = snapshot_download("DAVIAN-Robotics/pi05-robocasa-H50")
policy = get_policy_class("pi05").from_pretrained(path).to("cuda").eval()
pre, post = make_pre_post_processors(policy_cfg=policy.config, pretrained_path=path)
pre / post ๋ฅผ ์ง์ ๋ง๋ค๋ฉด ์ ๊ทํ ํต๊ณ๊ฐ ์ด๊ธ๋ ์ฑ๊ณต๋ฅ ์ด ์กฐ์ฉํ 0 ์ผ๋ก ๋จ์ด์ง๋ค โ
๋ฐ๋์ pretrained_path= ๋ก ์ฒดํฌํฌ์ธํธ์ baked ๋ ํต๊ณ๋ฅผ ๋ก๋ํ๋ผ.
์ตํฐ๋ง์ด์ ์ํ(training_state/, 13.4GB)๋ ํฌํจํ์ง ์์๋ค. ์ถ๋ก /ํ๊ฐ ์ ์ฉ์ด๋ค.
์ถ์ฒ ยท ๋ผ์ด์ ์ค
- ํ์ต ์ฝ๋: LeRobot (Apache-2.0), base ๋ชจ๋ธ
lerobot/pi05_base(Physical Intelligence ฯ0.5 ๋ฅผ LeRobot ์ด ํฌํ ). - ์๋ฎฌ๋ ์ดํฐ: RoboCasa / robosuite (MIT).
- ๋ณธ ์ฒดํฌํฌ์ธํธ: Apache-2.0. base ๋ชจ๋ธ๊ณผ ๋ฐ์ดํฐ์ ์ ์ ๋ผ์ด์ ์ค/์ด์ฉ ์ฝ๊ด์ ํจ๊ป ๋ฐ๋ฅธ๋ค.
- ์ ์: DAVIAN Lab (KAIST) โ ๊ฐ์
sk-260713-vision-lecture์ค์ต์ฉ.
Tokenizer
tokenizer/ contains the PaliGemma tokenizer (google/paligemma-3b-pt-224), redistributed so that
the policy can be loaded without HuggingFace authentication. The upstream repository is gated;
policy_preprocessor.json references it by name, so AutoTokenizer.from_pretrained() would
otherwise fail with a 401 for users who have not accepted Google's terms.
Gemma is provided under and subject to the Gemma Terms of Use. By using these tokenizer files you agree to those terms. See also the Gemma Prohibited Use Policy.
- Downloads last month
- 55
Model tree for DAVIAN-Robotics/pi05-robocasa-H50
Base model
lerobot/pi05_base