模型下载与依赖准备¶

📦 本页整理 ACT、SmolVLA、Pi0、Pi0.5、XVLA 的离线模型依赖准备方式。

⚠️ 先看冲突关系¶

Pi0/Pi0.5 与 SmolVLA/XVLA 的 transformers 依赖冲突
建议按模型族拆分 conda 环境

推荐环境拆分：

lerobot-act：ACT
lerobot-vla：SmolVLA / XVLA
lerobot-pi：Pi0 / Pi0.5

1. ACT¶

mkdir -p ~/.cache/torch/hub/checkpoints
wget https://download.pytorch.org/models/resnet18-f37072fd.pth -O ~/.cache/torch/hub/checkpoints/resnet18-f37072fd.pth

2. SmolVLA¶

mkdir -p ~/.cache/huggingface/hub/models--lerobot--smolvla_base
mkdir -p ~/.cache/huggingface/hub/models--HuggingFaceTB--SmolVLM2-500M-Video-Instruct

cp -r "/media/openarm/新加卷/weilindong/888/训练模型/smolvla_base"/* \
  ~/.cache/huggingface/hub/models--lerobot--smolvla_base/

cp -r "/media/openarm/新加卷/weilindong/888/训练模型/smolvla_模型依赖/SmolVLM2-500M-Video-Instruct"/* \
  ~/.cache/huggingface/hub/models--HuggingFaceTB--SmolVLM2-500M-Video-Instruct/

lerobot-env
pip install 'transformers>=4.57.1' 'num2words>=0.5.14' 'accelerate>=1.7.0' 'safetensors>=0.4.3' \
  -i https://mirrors.aliyun.com/pypi/simple/

3. Pi0¶

mkdir -p ~/.cache/huggingface/hub/models--lerobot--pi0_base
mkdir -p ~/.cache/huggingface/hub/models--google--paligemma-3b-pt-224
mkdir -p ~/.cache/huggingface/hub/models--physical-intelligence--fast

cp -r "/media/openarm/新加卷/weilindong/888/训练模型/pi0_base"/* \
  ~/.cache/huggingface/hub/models--lerobot--pi0_base/

cp -r "/media/openarm/新加卷/weilindong/888/训练模型/pi0_pi0fast_视觉依赖/paligemma-3b-pt-224"/* \
  ~/.cache/huggingface/hub/models--google--paligemma-3b-pt-224/

cp -r "/media/openarm/新加卷/weilindong/888/训练模型/Pi0_Pi0fast_WALLX_动作编码依赖/physical-intelligence-fast"/* \
  ~/.cache/huggingface/hub/models--physical-intelligence--fast/

lerobot-env
pip install "transformers @ git+https://ghp.ci/https://github.com/huggingface/transformers.git@fix/lerobot_openpi"

4. Pi0.5¶

mkdir -p ~/.cache/huggingface/hub/models--lerobot--pi05_base
mkdir -p ~/.cache/huggingface/hub/models--google--paligemma-3b-pt-224
mkdir -p ~/.cache/huggingface/hub/models--physical-intelligence--fast

cp -r "/media/openarm/新加卷/weilindong/888/训练模型/pi05_base"/* \
  ~/.cache/huggingface/hub/models--lerobot--pi05_base/

cp -r "/media/openarm/新加卷/weilindong/888/训练模型/pi0_pi0fast_视觉依赖/paligemma-3b-pt-224"/* \
  ~/.cache/huggingface/hub/models--google--paligemma-3b-pt-224/

cp -r "/media/openarm/新加卷/weilindong/888/训练模型/Pi0_Pi0fast_WALLX_动作编码依赖/physical-intelligence-fast"/* \
  ~/.cache/huggingface/hub/models--physical-intelligence--fast/

lerobot-env
pip install "transformers @ git+https://ghp.ci/https://github.com/huggingface/transformers.git@fix/lerobot_openpi"

5. XVLA（LeRobot 0.4.3）¶

mkdir -p ~/.cache/huggingface/hub/models--lerobot--xvla-base
mkdir -p ~/.cache/huggingface/hub/models--facebook--bart-large

cp -r "/media/openarm/新加卷/weilindong/888/训练模型/xvla-base"/* \
  ~/.cache/huggingface/hub/models--lerobot--xvla-base/

cp -r "/media/openarm/新加卷/weilindong/888/训练模型/xvla_模型依赖/bart-large"/* \
  ~/.cache/huggingface/hub/models--facebook--bart-large/

lerobot-env
pip install 'transformers>=4.57.1' -i https://mirrors.aliyun.com/pypi/simple/

✅ 验证建议¶

检查缓存目录是否有 config.json、model.safetensors 等核心文件
在目标环境执行一次 lerobot-train --help，确认依赖可导入