Model Training模型训练

10. Model training10. 模型训练

Run all commands below from ../lerobot. Before training, replace YOUR_NAME/YOUR_DATASET and YOUR_RUN, and confirm that the new output directory does not already exist so that new results remain separate from previous training runs.以下命令都从 ../lerobot 运行。开始训练前需要修改 YOUR_NAME/YOUR_DATASET 和 YOUR_RUN，并确认新的输出目录不存在，避免与以前的训练结果混在一起。

10.1 ACT training10.1 ACT 训练

ACT can directly read the current bimanual dataset's 12-dimensional states, 12-dimensional actions and three original camera keys. Camera keys do not need to be renamed.ACT 可以直接读取当前双臂数据集中的 12 维状态、12 维动作以及三个原始相机键，不需要进行相机键重命名。

HF_HUB_DISABLE_XET=1 lerobot-train \
  --dataset.repo_id=YOUR_NAME/YOUR_DATASET \
  --dataset.video_backend=pyav \
  --policy.type=act \
  --policy.device=cuda \
  --policy.use_amp=true \
  --policy.push_to_hub=false \
  --output_dir=outputs/train/ACT/YOUR_RUN \
  --job_name=YOUR_RUN \
  --batch_size=8 \
  --num_workers=4 \
  --steps=20000 \
  --save_checkpoint=true \
  --save_freq=5000 \
  --log_freq=100 \
  --tolerance_s=0.05 \
  --wandb.enable=false

These are the main settings used by the existing models in outputs/train/ACT/: AMP, batch size 8, 4 data-loading workers, 20000 training steps and a checkpoint every 5000 steps. ACT uses 100-step action chunks by default. If you change this, also check the data length and the action-execution settings used for real-robot evaluation.这与当前 outputs/train/ACT/ 中已有模型使用的主要参数一致：AMP、batch size 8、4 个数据读取进程、训练 20000 步、每 5000 步保存一次。ACT 默认使用 100 步 action chunk；如需改变，必须同时确认数据长度和实机评估时的动作执行设置。

If the dataset is outside Hugging Face's default LeRobot cache location, you can also add:如果数据集不在 Hugging Face 的默认 LeRobot 缓存位置，可以额外增加：

--dataset.root=/实际的数据集目录

10.2 SmolVLA training10.2 SmolVLA 训练

The existing task01–task08 models were fine-tuned from the RoboTwin-pretrained models/smolvla_robotwin. The following command follows the train_config.json files saved with the existing models.现有 task01～task08 模型是从 RoboTwin 预训练的 models/smolvla_robotwin 继续微调。以下命令按照现有模型保存的 train_config.json 整理。

HF_HUB_DISABLE_XET=1 \
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
lerobot-train \
  --config_path=BREE/configs/smolvla_train.json \
  --dataset.repo_id=YOUR_NAME/YOUR_DATASET \
  --dataset.video_backend=pyav \
  --policy.path=models/smolvla_robotwin \
  --policy.device=cuda \
  --policy.use_amp=true \
  --policy.push_to_hub=false \
  --policy.chunk_size=50 \
  --policy.n_action_steps=50 \
  --policy.max_state_dim=32 \
  --policy.max_action_dim=32 \
  --policy.resize_imgs_with_padding='[512, 512]' \
  --policy.freeze_vision_encoder=true \
  --policy.train_expert_only=true \
  --policy.train_state_proj=true \
  --policy.optimizer_lr=0.0001 \
  --policy.optimizer_weight_decay=1e-10 \
  --policy.scheduler_warmup_steps=1000 \
  --policy.scheduler_decay_steps=20000 \
  --policy.scheduler_decay_lr=2.5e-6 \
  --batch_size=12 \
  --num_workers=4 \
  --steps=30000 \
  --save_checkpoint=true \
  --save_freq=5000 \
  --log_freq=500 \
  --tolerance_s=0.05 \
  --wandb.enable=false

The existing SmolVLA task models share these settings: three cameras, 12-dimensional real actions, 50-step action chunks, maximum state and action dimensions of 32, images padded to 512×512, a frozen vision encoder, and training of the expert and state projection. The three-camera rename_map is already defined in BREE/configs/smolvla_train.json.现有 SmolVLA 任务模型的共同配置为：三个相机、12 维真实动作、50 步 action chunk、状态和动作最大维度 32、图像 padding 到 512×512、冻结视觉编码器，并训练 expert 与 state projection。三个相机的 rename_map 已写在 BREE/configs/smolvla_train.json 中。

You must use --config_path here: the command-line configuration parser used in this repository cannot correctly parse the top-level --rename_map dictionary or the optional path type of --output_dir; passing them directly on the command line causes parameter-type errors. Before training, replace YOUR_NAME/YOUR_DATASET in the command with the actual dataset, then open BREE/configs/smolvla_train.json and replace both occurrences of YOUR_RUN with the training-run name. The command-line --dataset.repo_id overrides the dataset placeholder in the configuration template.这里必须使用 --config_path：当前仓库所用的命令行配置解析器不能正确解析顶层 --rename_map 字典以及可选路径类型的 --output_dir，直接在命令中传入会报参数类型错误。训练前需要修改两处：将命令中的 YOUR_NAME/YOUR_DATASET 换成真实数据集；将 BREE/configs/smolvla_train.json 中两处 YOUR_RUN 换成训练名称。命令中的 --dataset.repo_id 会覆盖配置模板里的数据集占位值。

10.3 XVLA training10.3 XVLA 训练

The XVLA base model uses image, image2 and image3 as its three visual-input keys, so the current dataset's global, left-wrist and right-wrist camera keys must be renamed. action_mode=auto automatically identifies the bimanual dataset's 12-dimensional real actions and applies padding and trimming within XVLA's 20-dimensional model action space.XVLA 基础模型使用 image、image2、image3 作为三个视觉输入键，因此必须把当前数据集的全局、左腕和右腕相机键重命名。action_mode=auto 会根据双臂数据集自动识别真实的 12 维动作，并在 XVLA 的 20 维模型动作空间中完成填充和裁剪。

The settings below follow existing XVLA training results on this machine: 3 cameras, 30-step action chunks, bfloat16, batch size 2, and training of the policy transformer and soft prompts.下面的参数根据本机已有 XVLA 训练结果整理：使用 3 个相机、30 步 action chunk、bfloat16、batch size 2，并训练 policy transformer 和 soft prompts。

HF_HUB_DISABLE_XET=1 \
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
lerobot-train \
  --dataset.repo_id=YOUR_NAME/YOUR_DATASET \
  --dataset.video_backend=pyav \
  --policy.path=lerobot/xvla-base \
  --policy.device=cuda \
  --policy.use_amp=true \
  --policy.push_to_hub=false \
  --policy.dtype=bfloat16 \
  --policy.action_mode=auto \
  --policy.max_state_dim=20 \
  --policy.max_action_dim=20 \
  --policy.chunk_size=30 \
  --policy.n_action_steps=30 \
  --policy.resize_imgs_with_padding='[224, 224]' \
  --policy.num_image_views=3 \
  --policy.freeze_vision_encoder=true \
  --policy.freeze_language_encoder=true \
  --policy.train_policy_transformer=true \
  --policy.train_soft_prompts=true \
  --rename_map='{
    "observation.images.left_global": "observation.images.image",
    "observation.images.left_wrist": "observation.images.image2",
    "observation.images.right_wrist": "observation.images.image3"
  }' \
  --output_dir=outputs/train/XVLA/YOUR_RUN \
  --job_name=YOUR_RUN \
  --batch_size=2 \
  --num_workers=2 \
  --steps=30000 \
  --save_checkpoint=true \
  --save_freq=5000 \
  --log_freq=500 \
  --tolerance_s=0.05 \
  --wandb.enable=false

The current command freezes XVLA's vision and language encoders and trains only the policy Transformer and soft prompts, keeping GPU memory use relatively low. To fine-tune the full VLM, set both freeze parameters to false, but expect substantially higher memory use and training time.当前命令冻结 XVLA 的视觉和语言编码器，只训练策略 Transformer 和 soft prompts，显存占用相对较低。如果需要完整微调 VLM，可以把两个 freeze 参数改成 false，但显存和训练时间都会明显增加。

Before the first training run, confirm that the XVLA extra dependencies are installed and that lerobot/xvla-base can be downloaded from Hugging Face or is already cached locally. The base model on the current machine is cached at:首次训练前确认已安装 XVLA 额外依赖，并且 lerobot/xvla-base 可以从 Hugging Face 下载或已经存在于本机缓存中。当前机器的基础模型已经缓存在：

~/.cache/huggingface/hub/models--lerobot--xvla-base/

If all models and data are cached and you need fully offline training, you can also set the following at the start of the command:如果确认所有模型和数据都已缓存，需要完全离线训练时，可以在命令开头额外设置：

HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1

10.4 RDT and OpenPI training10.4 RDT 与 OpenPI 训练

The RDT and OpenPI training functionality is organized under:RDT 和 OpenPI 训练功能已经整理到：

BREE/model_training/

By default, training scripts read their dependencies from models/ and datasets/ in the current LeRobot repository and write checkpoints to outputs/train/.训练脚本默认从当前 LeRobot 仓库下的 models/ 和 datasets/ 读取依赖，并将 checkpoint 写入 outputs/train/。

10.4.1 OpenPI training10.4.1 OpenPI 训练

OpenPI reads data from datasets/openpi/ and base weights from models/openpi/pi05_base/params/ by default, and saves checkpoints to outputs/train/OpenPI/. For the first run, start with a 1-step smoke test:OpenPI 默认读取 datasets/openpi/ 下的数据和 models/openpi/pi05_base/params/ 下的基础权重，并把 checkpoint 保存到 outputs/train/OpenPI/。第一次训练建议先执行 1 步 smoke test：

GPU_IDS=0,1 \
BREE/model_training/openpi/train_openpi.sh pi05_biso_task01 smoke

After confirming that the data, weights, GPUs, FSDP and checkpoint writing work correctly, start a full training run:确认数据、权重、GPU、FSDP 和 checkpoint 写入都正常后，再开始正式训练：

GPU_IDS=0,1 \
BREE/model_training/openpi/train_openpi.sh pi05_biso_task01 train

To resume the same experiment from an existing checkpoint, EXP_NAME must match the original experiment name:需要从已有 checkpoint 恢复同一个实验时，EXP_NAME 必须与原实验名称一致：

GPU_IDS=0,1 RESUME=1 EXP_NAME=my_test \
BREE/model_training/openpi/train_openpi.sh pi05_biso_task03_switch_new train

If you currently have only the HDF5 data used by RDT, first convert it to the LeRobot v2.1 format used by OpenPI, then compute normalization statistics:如果当前只有 RDT 使用的 HDF5 数据，先转换成 OpenPI 使用的 LeRobot v2.1 数据，再计算归一化统计：

BREE/model_training/openpi/convert_from_rdt_hdf5.sh \
  datasets/rdt_hdf5/task01_pickplace_big_50 \
  yyyyyy0604/task01_pickplace_big_50

BREE/model_training/openpi/compute_norm_stats.sh pi05_biso_task01

Here, GPU_IDS selects the physical GPUs used for training; the first positional argument is the OpenPI training configuration name; smoke trains for only 1 step, while train starts full training; and RESUME=1 resumes training. Standard tasks train for 30000 steps by default, while some long-prompt, wrist-swapped and switch variants default to 20000 steps. Override this with TRAIN_STEPS=步数.其中 GPU_IDS 指定参与训练的物理 GPU；第一个位置参数是 OpenPI 训练配置名；smoke 只训练 1 步，train 执行正式训练；RESUME=1 表示恢复训练。标准任务默认训练 30000 步，部分 long-prompt、wrist-swapped 和 switch 变体默认训练 20000 步，可用 TRAIN_STEPS=步数 覆盖。

10.4.2 RDT training10.4.2 RDT 训练

RDT reads data from datasets/rdt_hdf5/ and the RDT, SigLIP and T5 weights from models/rdt/ by default, and saves checkpoints to outputs/train/RDT/. If you currently have only LeRobot data, convert it first:RDT 默认读取 datasets/rdt_hdf5/ 下的数据和 models/rdt/ 下的 RDT、SigLIP、T5 权重，并把 checkpoint 保存到 outputs/train/RDT/。如果目前只有 LeRobot 数据，先执行转换：

BREE/model_training/rdt/convert_dataset.sh \
  yyyyyy0604/task08_cylinder_insertion_50

Then start a full training run. RDT_PYTHON must point to a Python interpreter with the RDT dependencies installed. You can omit this variable if those dependencies are already installed correctly in the current environment:然后执行正式训练。RDT_PYTHON 要指向已经安装 RDT 依赖的 Python；如果当前环境已经正确安装依赖，也可以省略这个变量：

GPU=0 \
RDT_PYTHON=/实际的/rdt环境/bin/python \
BREE/model_training/rdt/train_rdt.sh \
  task08_insertion_170m \
  task08_cylinder_insertion_50 \
  "Pick up the red cylinder with the right arm, then place the blue cylinder onto the red cylinder with the left arm"

Resume an existing training run:恢复已有训练：

GPU=0 RESUME=1 \
RDT_PYTHON=/实际的/rdt环境/bin/python \
BREE/model_training/rdt/train_rdt.sh \
  task08_insertion_170m \
  task08_cylinder_insertion_50 \
  "Pick up the red cylinder with the right arm, then place the blue cylinder onto the red cylinder with the left arm"

The three positional arguments are, in order: the task profile, the data-directory name under datasets/rdt_hdf5/, and the task's language instruction. GPU selects the physical GPU, and RESUME=1 resumes from an existing checkpoint in the output directory. By default, training runs for 25000 steps and saves a checkpoint every 5000 steps. Change these with MAX_TRAIN_STEPS and CHECKPOINT_PERIOD, respectively.三个位置参数依次为：任务 profile、datasets/rdt_hdf5/ 下的数据目录名、任务语言指令。GPU 指定物理 GPU，RESUME=1 表示从输出目录中的已有 checkpoint 恢复。默认训练 25000 步、每 5000 步保存一次，可分别用 MAX_TRAIN_STEPS 和 CHECKPOINT_PERIOD 修改。

The training code has been merged into BREE, but the code directory does not include the base models, training data or RDT Python environment. Before training, place them in the default directories above or use the environment variables supported by the training scripts to point to their actual locations.训练代码已经合并进 BREE，但基础模型、训练数据和 RDT Python 环境没有包含在代码目录中；开始训练前必须把它们放入上述默认目录，或使用训练脚本支持的环境变量指向实际位置。

For the complete list of configuration names, RDT profiles, installation dependencies and overridable paths, see:完整的配置名称、RDT profile、安装依赖以及所有可覆盖路径见：

BREE/model_training/README.md

The complete contents of BREE/model_training/README.md follow.下方完整展开 BREE/model_training/README.md 的内容。

BREE model trainingBREE 模型训练

This directory contains only the RDT and OpenPI training entry points, task configurations and data-conversion tools. Runtime files, language embeddings and checkpoints generated during training are written to outputs/ at the repository root, rather than to BREE.本目录只保存 RDT 和 OpenPI 的训练入口、任务配置与数据转换工具。训练产生的 runtime、语言向量和 checkpoint 统一写入仓库根目录的 outputs/，不会写入 BREE。

All launch scripts locate the /home/l/lerobot repository root from their own location and do not depend on the original server's /home/yanch/... paths. Environment variables can override the default locations for models or data stored on external disks.所有启动脚本都会根据自身位置查找 /home/l/lerobot 仓库根目录，不依赖原服务器的 /home/yanch/...。外部磁盘上的模型或数据可以通过环境变量覆盖默认位置。

1. OpenPI training1. OpenPI 训练

OpenPI training reuses:OpenPI 训练复用：

BREE/openpi_eval/openpi/

Both the downloaded code and the existing OpenPI code use commit 215abfb217dbac7d5f1273282331b9b1866c0479. The following have been merged:下载代码和现有 OpenPI 的 commit 都是 215abfb217dbac7d5f1273282331b9b1866c0479。已经合并：

• 12 BiSO training configurations;12 个 BiSO 训练配置；
• 12 norm_stats.json files;12 份 norm_stats.json；
• A fix for LeRobot video_backend loading;LeRobot video_backend 加载修复；
• HDF5-to-OpenPI dataset conversion tools;HDF5 到 OpenPI 数据集转换工具；
• Unified entry points for smoke tests, full training and resumed training.统一的 smoke test、正式训练和恢复训练入口。

1.1 Default directories1.1 默认目录

models/openpi/pi05_base/params/       OpenPI 基础权重
datasets/openpi/<用户>/<数据集>/       OpenPI/LeRobot v2.1 训练数据
outputs/train/OpenPI/                 新训练 checkpoint

Execution is offline by default; missing models or data will not be downloaded automatically.默认离线运行，不会自动下载缺失的模型或数据。

1.2 Supported configurations1.2 支持的配置

pi05_biso_task01
pi05_biso_task02
pi05_biso_task02_long_prompt
pi05_biso_task02_long_prompt_wrist_swapped
pi05_biso_task03
pi05_biso_task03_switch_new
pi05_biso_task03_switch_right
pi05_biso_task04
pi05_biso_task05
pi05_biso_task06
pi05_biso_task07
pi05_biso_task08

1.3 Smoke test

The smoke test trains for only 1 step to verify the data, weights, GPUs, FSDP and checkpoint writing:Smoke test 只训练 1 步，用于验证数据、权重、GPU、FSDP 和 checkpoint 写入：

GPU_IDS=0,1 \
BREE/model_training/openpi/train_openpi.sh pi05_biso_task01 smoke

1.4 Full training1.4 正式训练

GPU_IDS=0,1 \
BREE/model_training/openpi/train_openpi.sh pi05_biso_task01 train

Standard tasks train for 30000 steps by default; long-prompt, wrist-swapped, switch-new and switch-right variants default to 20000 steps. Override this with TRAIN_STEPS:标准任务默认训练 30000 步；long-prompt、wrist-swapped、switch-new 和 switch-right 默认训练 20000 步。可用 TRAIN_STEPS 覆盖：

GPU_IDS=0,1 TRAIN_STEPS=10000 EXP_NAME=my_test \
BREE/model_training/openpi/train_openpi.sh pi05_biso_task03_switch_new train

Resume an existing training run:恢复已有训练：

GPU_IDS=0,1 RESUME=1 EXP_NAME=my_test \
BREE/model_training/openpi/train_openpi.sh pi05_biso_task03_switch_new train

1.5 OpenPI data conversion and statistics1.5 OpenPI 数据转换和统计

Convert existing RDT HDF5 data to the LeRobot v2.1 format used by OpenPI:将已有 RDT HDF5 转为 OpenPI 使用的 LeRobot v2.1 数据：

BREE/model_training/openpi/convert_from_rdt_hdf5.sh \
  datasets/rdt_hdf5/task01_pickplace_big_50 \
  yyyyyy0604/task01_pickplace_big_50

Compute normalization statistics for a new dataset:为新数据集计算归一化统计：

BREE/model_training/openpi/compute_norm_stats.sh pi05_biso_task01

1.6 OpenPI path overrides1.6 OpenPI 可覆盖路径

2. RDT training2. RDT 训练

RDT reuses the shared source code already in the repository:RDT 复用仓库已有公共源码：

third_party/RoboticsDiffusionTransformer/

The shared code in all 13 runtime sets in the download package is identical, so only task-specific configurations are retained. Before training, an isolated working copy is created at outputs/rdt_runtime/<profile>/, leaving the shared source code unchanged.下载包中的 13 套 runtime 公共代码完全相同，因此只保留各任务独有的配置。训练前会在 outputs/rdt_runtime/<profile>/ 创建隔离工作副本，不会覆盖公共源码。

2.1 Default directories2.1 默认目录

models/rdt/rdt-170m/                  RDT-170M 基础权重
models/rdt/rdt-1b/                    RDT-1B 基础权重
models/rdt/siglip-so400m-patch14-384/ 视觉编码器
models/rdt/t5-v1_1-xxl/               T5 文本编码器
datasets/lerobot/<用户>/<数据集>/       原始 LeRobot 数据
datasets/rdt_hdf5/<任务名>/            转换后的 RDT HDF5 数据
outputs/rdt_runtime/                  自动生成的隔离 runtime
outputs/rdt_t5_offload/               T5 编码临时文件
outputs/train/RDT/                    新训练 checkpoint

2.2 Data conversion2.2 数据转换

BREE/model_training/rdt/convert_dataset.sh \
  yyyyyy0604/task08_cylinder_insertion_50

If the original data is outside the default location, pass its actual directory as the second argument:如果原始数据不在默认位置，可以把实际目录作为第二个参数：

BREE/model_training/rdt/convert_dataset.sh \
  yyyyyy0604/task08_cylinder_insertion_50 \
  /实际的/LeRobot数据目录

2.3 Full training2.3 正式训练

Before running, activate the RDT environment or use RDT_PYTHON to specify its Python interpreter:运行前需要激活 RDT 环境，或者通过 RDT_PYTHON 指定该环境的 Python：

GPU=0 \
RDT_PYTHON=/实际的/rdt环境/bin/python \
BREE/model_training/rdt/train_rdt.sh \
  task08_insertion_170m \
  task08_cylinder_insertion_50 \
  "Pick up the red cylinder with the right arm, then place the blue cylinder onto the red cylinder with the left arm"

The default parameters match the server's standard 170M training setup: 25000 steps, batch size 2, gradient accumulation of 4, global batch size 8, learning rate 1e-4, and a checkpoint every 5000 steps.默认参数与服务器标准 170M 训练一致：25000 步、batch size 2、梯度累积 4、全局 batch 8、学习率 1e-4、每 5000 步保存。

Resume training:恢复训练：

GPU=0 RESUME=1 \
BREE/model_training/rdt/train_rdt.sh PROFILE TASK_NAME "INSTRUCTION"

2.4 RDT profile

Available profiles are located in BREE/model_training/rdt/task_profiles/:可用 profile 位于 BREE/model_training/rdt/task_profiles/：

task01_02_170m
task02_long_prompt_170m
task02_long_prompt_wrist_swapped_170m
task03_170m
task03_switch_new_170m
task03_switch_right_170m
task04_1b
task04_1b_visual_lr1e-5
task04_masked_loss_repair_170m
task06_170m
task07_08_170m
task08_insertion_170m
task10_170m

The long-prompt, wrist-swapped, switch-new, switch-right and task04 masked-loss profiles use masked loss, so only the real robot action dimensions contribute to the loss; all other profiles use standard MSE. Profiles with 1b in their names load rdt-1b by default; the others load rdt-170m.其中 long-prompt、wrist-swapped、switch-new、switch-right 和 task04 masked-loss profile 使用 masked loss，只让真实机器人动作维度参与 loss；其余 profile 使用标准 MSE。名称包含 1b 的 profile 默认加载 rdt-1b，其他默认加载 rdt-170m。

2.5 RDT parameter overrides2.5 RDT 可覆盖参数

3. Files not included3. 未包含的文件

The server download package did not include the following, so this integration does not present them as already available:服务器下载包没有携带以下内容，因此本次也没有把它们伪装成已经可用：

• RDT-170M, RDT-1B, SigLIP and T5 weights;RDT-170M、RDT-1B、SigLIP、T5 权重；
• OpenPI π0.5 base weights;OpenPI π0.5 基础权重；
• RDT/OpenPI training data;RDT/OpenPI 训练数据；
• The RDT Conda environment;RDT Conda 环境；
• Historical checkpoints from the server.服务器上的历史 checkpoint。

The training scripts can run only after these files have been placed in the default directories above or their actual locations have been specified through environment variables.把这些内容放入上述默认目录，或通过环境变量指向实际位置后，训练脚本才可以运行。