VLA模型

# VLA-models 本仓库存储了基于 [LeRobot](https://github.com/huggingface/lerobot) 框架训练的机器人操作策略模型权重，任务为单臂桌面抓取放置（将杯子抓起放入盘中），使用 SO-101 机械臂采集，共 100 个演示轨迹。 --- ## 目录结构与说明 ### $π_{0.5}$ (pi0.5) — 只训练动作专家头，冻结视觉编码器 | 路径 | 数据量 | 训练步数 | 说明 | | ----------------------------- | ------------ | ----------- | -------------------------------- | | `pi05_10k/009000/` | 100 episodes | 9000 steps | 倒数第二个checkpoint | | `pi05_10k/010000/` | 100 episodes | 10000 steps | 最终checkpoint | | `pi05_50ep/009000/` | 50 episodes | 9000 steps | 数据量消融，倒数第二个checkpoint | | `pi05_50ep/010000/` | 50 episodes | 10000 steps | 数据量消融，最终checkpoint | | `pi05_training_small/001500/` | 100 episodes | 1500 steps | 短训练消融，最终checkpoint | | `pi05_10k_64bs/001250/` | 100 episodes | 1250 steps | batch_size=64 | | `pi05_50ep_64bs/001250` | 50 episodes | 1250 steps | batch_size=64 | ### ACT — 全量训练，100k steps，save_freq=20k | 路径 | 说明 | | ---------------------- | ---------------------------------------------------- | | `act_demo_all/{step}/` | 基准实验，100 episodes | | `act_50ep/{step}/` | 数据量消融，50 episodes | | `act_augment/{step}/` | 开启图像数据增强，100 episodes | | `act_ensemble/{step}/` | 开启 temporal ensemble（coeff=0.01），100 episodes | | `act_large/{step}/` | 扩大模型（dim_model=1024, n_heads=16），100 episodes | | `act_64bs/012500/` | 100 episode，batch_size=64 | | `act_200k/*/` | 100 episode，200k/240k/320k/400k steps，batch_size=64 | | `act_last_50/012500/` | 后 50 episode，batch_size=64 | | `act_50ep_64bs/012500/` | 50 episode，batch_size=64 | 其中 `{step}` 为 `020000` / `040000` / `060000` / `080000` / `100000`。 ### Diffusion — 全量训练，100k steps | 路径 | 说明 | | ---------- | -------------------------------------- | | `dp_demo/` | 100 episodes，batch_size=8 | | `dp_50ep/` | 50 episodes，batch_size=32，step=25000 | --- ## 下载方式 ```bash modelscope download --dataset 'JuShenRobo/VLA-models' --include 'act_demo_all/100000/*' ``` ## 加载模型 ```python from lerobot.policies.act.modeling_act import ACTPolicy from lerobot.policies.pi05.modeling_pi05 import PI05Policy # 加载 ACT policy = ACTPolicy.from_pretrained('本地路径/act_demo_all/100000/') # 加载 pi0.5 policy = PI05Policy.from_pretrained('本地路径/pi05_10k/010000/') ```