pi0.5 โ€” RoboCasa H50 (fine-tuned)

๊ฐ•์˜ sk-260713-vision-lecture (SK ๋น„์ „ ๊ฐ•์˜, ์‹œ๋ฎฌ๋ ˆ์ด์…˜ VLA ์‹ค์Šต) ์˜ baseline ์ฒดํฌํฌ์ธํŠธ๋‹ค. LeRobot ์˜ pi05 ์ •์ฑ…(ฯ€0.5, flow-matching VLA)์„ lerobot/pi05_base ์—์„œ ์ถœ๋ฐœํ•ด RoboCasa ํ‚ค์นœ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ full fine-tuning ํ–ˆ๋‹ค.

๋ฐ์ดํ„ฐ

  • DAVIAN-Robotics/robocasa-H50 โ€” 1,293 ์—ํ”ผ์†Œ๋“œ / 242 ํƒœ์Šคํฌ (LeRobot v3.0 ํฌ๋งท, RoboCasa ํ‚ค์นœ).

ํ•™์Šต ์„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€ ํ•™์Šต)
chunk size 50
dtype bfloat16
optimizer AdamW (betas 0.9/0.95, wd 0.01, grad clip 1.0)

๋ชจ๋ธ config (config.json)

  • chunk_size = 50, n_action_steps = 25, num_inference_steps = 10 (flow matching ์ ๋ถ„ ์Šคํ…)
  • ์นด๋ฉ”๋ผ 3๋Œ€ โ€” robot0_agentview_left_image, robot0_agentview_right_image, robot0_eye_in_hand_image. ๊ฐ 3ร—128ร—128 ์ž…๋ ฅ์ด๋ฉฐ ์ •์ฑ… ๋‚ด๋ถ€์—์„œ 224ร—224 (image_resolution) ๋กœ ๋ฆฌ์‚ฌ์ด์ฆˆ๋œ๋‹ค.
  • state 16-d (max_state_dim=32 ๋กœ ํŒจ๋”ฉ), action 12-d (max_action_dim=32 ๋กœ ํŒจ๋”ฉ)
  • ์ •๊ทœํ™”: VISUAL=IDENTITY, STATE/ACTION=QUANTILES โ€” ํ†ต๊ณ„๋Š” policy_preprocessor_*.safetensors / policy_postprocessor_*.safetensors ์— baked ๋˜์–ด ์žˆ๋‹ค. ๋ฐ˜๋“œ์‹œ ํ•จ๊ป˜ ๋กœ๋“œํ•ด์•ผ ํ•œ๋‹ค (์•„๋ž˜ ์ฐธ์กฐ).

๊ฒ€์ฆ๋œ ์„ฑ๋Šฅ

PnPCounterToCab, 40 ์—ํ”ผ์†Œ๋“œ, max_steps=500:

์กฐ๊ฑด Success Rate
์›๋ณธ ๋žœ๋ค ์”ฌ (layout/style/์˜ค๋ธŒ์ ํŠธ ์ „๋ถ€ ๋žœ๋ค โ€” ํ•™์Šต ๋ถ„ํฌ์™€ ๋™์ผ) 0.550
๊ณ ์ •-apple ์”ฌ (layout 0 / style 0 / apple โ€” ๊ฐ•์˜ ์‹ค์Šต ๊ธฐ๋ณธ config) 0.575

ฯ€0.5 ๋Š” flow matching ์ด๋ผ ์ฒญํฌ๋งˆ๋‹ค ๋…ธ์ด์ฆˆ๋ฅผ ์ƒˆ๋กœ ๋ฝ‘๋Š”๋‹ค โ†’ ๊ฐ™์€ env seed ์—์„œ๋„ SR ์ด ํ”๋“ค๋ฆฐ๋‹ค. 40 ์—ํ”ผ์†Œ๋“œ ๊ธฐ์ค€ 95% Wilson CI ๋Š” ์•ฝ ยฑ0.15 ๋‹ค. ์ด ๋ฒ”์œ„ ๋ฐ–์˜ ์ฐจ์ด๋งŒ ์œ ์˜๋ฏธํ•˜๊ฒŒ ์ฝ์–ด๋ผ.

์‚ฌ์šฉ๋ฒ•

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 ์ง์ ‘ ๋งŒ๋“ค๋ฉด ์ •๊ทœํ™” ํ†ต๊ณ„๊ฐ€ ์–ด๊ธ‹๋‚˜ ์„ฑ๊ณต๋ฅ ์ด ์กฐ์šฉํžˆ 0 ์œผ๋กœ ๋–จ์–ด์ง„๋‹ค โ€” ๋ฐ˜๋“œ์‹œ pretrained_path= ๋กœ ์ฒดํฌํฌ์ธํŠธ์— baked ๋œ ํ†ต๊ณ„๋ฅผ ๋กœ๋“œํ•˜๋ผ.

์˜ตํ‹ฐ๋งˆ์ด์ € ์ƒํƒœ(training_state/, 13.4GB)๋Š” ํฌํ•จํ•˜์ง€ ์•Š์•˜๋‹ค. ์ถ”๋ก /ํ‰๊ฐ€ ์ „์šฉ์ด๋‹ค.

์ถœ์ฒ˜ ยท ๋ผ์ด์„ ์Šค

  • ํ•™์Šต ์ฝ”๋“œ: LeRobot (Apache-2.0), base ๋ชจ๋ธ lerobot/pi05_base (Physical Intelligence ฯ€0.5 ๋ฅผ LeRobot ์ด ํฌํŒ…).
  • ์‹œ๋ฎฌ๋ ˆ์ดํ„ฐ: RoboCasa / robosuite (MIT).
  • ๋ณธ ์ฒดํฌํฌ์ธํŠธ: Apache-2.0. base ๋ชจ๋ธ๊ณผ ๋ฐ์ดํ„ฐ์…‹์˜ ์› ๋ผ์ด์„ ์Šค/์ด์šฉ ์•ฝ๊ด€์„ ํ•จ๊ป˜ ๋”ฐ๋ฅธ๋‹ค.
  • ์ œ์ž‘: DAVIAN Lab (KAIST) โ€” ๊ฐ•์˜ sk-260713-vision-lecture ์‹ค์Šต์šฉ.

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
Safetensors
Model size
4B params
Tensor type
F32
ยท
BF16
ยท
Video Preview
loading

Model tree for DAVIAN-Robotics/pi05-robocasa-H50

Finetuned
(126)
this model

Dataset used to train DAVIAN-Robotics/pi05-robocasa-H50