LeRobot机械臂端到端全流程指南：从数据采集、云端训练到边缘推理

conda create -y -n lerobot python=3.10

# step:0 安装编译依赖
sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config
# step:1 激活环境
conda activate lerobot
# step:2 安装ffmpeg
conda install ffmpeg -c conda-forge

# 以下两种方式任选其一：
# step:3.1 安装方式一：从源码安装lerobot
git clone https://github.com/xiongqi123123/LeRobot-VLA.git
cd LeRobot-VLA/lerobot
pip install -e .
# step:3.2 安装方式二：从PyPI安装
pip install lerobot
# 要安装附加功能，请使用以下之一
pip install 'lerobot[all]'          # All available features
pip install 'lerobot[aloha,pusht]'  # Specific features (Aloha & Pusht)
pip install 'lerobot[feetech]'      # Feetech motor support

lerobot-find-port

lerobot-setup-motors \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem585A0076841  # <- paste here the port found at previous step

lerobot-calibrate \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --robot.id=my_awesome_follower_arm # <- Give the robot a unique name

lerobot-find-cameras opencv

lerobot-teleoperate \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem5AB90671801 \
    --robot.id=my_awesome_follower_arm \
    --teleop.type=so101_leader \
    --teleop.port=/dev/tty.usbmodem5AB90671501 \
    --teleop.id=my_awesome_leader_arm \

lerobot-teleoperate \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem5AB90671801 \
    --robot.id=my_awesome_follower_arm \
    --robot.cameras="{ arm: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}, front: {type: opencv, index_or_path: 2, width: 1920, height: 1080, fps: 30}}" \
    --teleop.type=so101_leader \
    --teleop.port=/dev/tty.usbmodem5AB90671501 \
    --teleop.id=my_awesome_leader_arm \
    --display_data=true

lerobot-record \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem5AB90671801 \
    --robot.id=my_awesome_follower_arm \
    --robot.cameras="{ arm: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}, front: {type: opencv, index_or_path: 2, width: 1920, height: 1080, fps: 30}}" \
    --teleop.type=so101_leader \
    --teleop.port=/dev/tty.usbmodem5AB90671501 \
    --teleop.id=my_awesome_leader_arm \
    --display_data=true \
    --dataset.repo_id=skyxz/blackmarker_scence1 \
    --dataset.num_episodes=5 \
    --dataset.single_task="Grab the black marker and put it in the bin" \
    --dataset.push_to_hub=False \
    --dataset.episode_time_s=15 \
    --dataset.reset_time_s=5 

aws configure
AWS Access Key ID [None]: ***********
AWS Secret Access Key [None]: **********************
Default region name [None]: ap-northeast-2
Default output format [None]:

pip3 install boto3 sagemaker

import boto3

boto_sess = boto3.Session(
    region_name="region"  
)

sm = boto_sess.client('sagemaker')
resp = sm.list_training_jobs(MaxResults=1)  
print("SageMaker Connected")

service_quotas = boto_sess.client('service-quotas')

quotas = service_quotas.list_service_quotas(ServiceCode='sagemaker')

print("=== SageMaker Instance Quota ===")
for q in quotas['Quotas']:
    if 'ml.' in q['QuotaName']:
        print(f"{q['QuotaName']} -> Limit: {q['Value']}")

import boto3
import json

boto_sess = boto3.Session(
    region_name="Region"  
)
iam = boto_sess.client("iam")
role_name = "SkyXZ-Test-Role"
trust_policy = {
    "Version": "2012-10-17",
    "Statement": [{
        "Effect": "Allow",
        "Principal": {"Service": "sagemaker.amazonaws.com"},
        "Action": "sts:AssumeRole"
    }]
}
try:
    role = iam.create_role(
        RoleName=role_name,
        AssumeRolePolicyDocument=json.dumps(trust_policy),
        Description="SageMaker execution role"
    )
except iam.exceptions.EntityAlreadyExistsException:
    role = iam.get_role(RoleName=role_name)
policies = [
    "arn:aws:iam::aws:policy/AmazonSageMakerFullAccess",
    "arn:aws:iam::aws:policy/AmazonS3FullAccess"
]
for p in policies:
    iam.attach_role_policy(RoleName=role_name, PolicyArn=p)

print(f"Role Created: {role['Role']['Arn']}")

import boto3
boto_sess = boto3.Session(
    region_name="Region"  
)
iam = boto_sess.client('iam')

paginator = iam.get_paginator('list_roles')
for page in paginator.paginate():
    for role in page['Roles']:
        print(f"Role Name: {role['RoleName']}")
        print(f"  ARN: {role['Arn']}")
        print(f"  Create Date: {role['CreateDate']}")
        print("-" * 60)

import boto3
import os
boto_sess = boto3.Session(
    region_name="Region"  
)
s3 = boto_sess.client('s3')
bucket_name = 'skyxz-test-bucket-2025'
s3.create_bucket(
    Bucket=bucket_name,
    CreateBucketConfiguration={'LocationConstraint': 'Region'}
)
print(f"S3 Bucket {bucket_name} Created")

import boto3
boto_sess = boto3.Session(
    region_name="Region"  
)
s3 = boto_sess.client('s3')
s3_res = boto_sess.resource('s3')

for bucket in s3.list_buckets()['Buckets']:
    name = bucket['Name']
    region = s3.get_bucket_location(Bucket=name).get('LocationConstraint')
    size_mb = sum(obj.size for obj in s3_res.Bucket(name).objects.all()) / (1024*1024)
    print(f"{name} | Region: {region} | Size: {size_mb:.2f} MB")

import boto3, os
boto_sess = boto3.Session(
    region_name="Region"  
)
s3 = boto_sess.client('s3')

bucket_name = 'your-bucket'
local_dir = './my_local_dataset'
s3_prefix = 'lerobot-data/' 

for root, dirs, files in os.walk(local_dir):
    for file in files:
        local_path = os.path.join(root, file)
        relative_path = os.path.relpath(local_path, local_dir)
        s3_key = os.path.join(s3_prefix, relative_path).replace("\\", "/") 
        s3.upload_file(local_path, bucket_name, s3_key)
        print(f"Success: {local_path} → s3://{bucket_name}/{s3_key}")

import boto3
boto_sess = boto3.Session(
    region_name="Region" 
)
s3 = boto_sess.client('s3')
bucket_name = "skyxz-test-bucket-2025"
s3_prefix = "training-data"
response = s3.list_objects_v2(Bucket=bucket_name, Prefix=s3_prefix)
if "Contents" in response:
    for obj in response["Contents"]:
        print(obj["Key"], obj["Size"], obj["LastModified"])
else:
    print("No Found Files")

import boto3
boto_sess = boto3.Session(
    region_name="Region" 
)
s3 = boto_sess.client('s3')
bucket_name = "skyxz-test-bucket-2025"
key_to_delete = "training-data/example.txt" 

s3.delete_object(Bucket=bucket_name, Key=key_to_delete)
print(f"{key_to_delete} have been {bucket_name} deleted")

import boto3
boto_sess = boto3.Session(
    region_name="Region"  
)
s3 = boto_sess.client('s3')
bucket_name = "skyxz-test-bucket-2025"
objects = s3.list_objects_v2(Bucket=bucket_name)
if 'Contents' in objects:
    for obj in objects['Contents']:
        s3.delete_object(Bucket=bucket_name, Key=obj['Key'])
s3.delete_bucket(Bucket=bucket_name)

print(f"{bucket_name} have been deleted")

lerobot-train \
     --dataset.repo_id=${HF_USER}/so101-test-0903 \
     --policy.type=act \
     --output_dir=outputs/train/act_so100_test \
     --job_name=act_so100_test \
     --policy.device=cuda \
     --wandb.enable=false \
     --policy.repo_id=${HF_USER}/my_policy

from sagemaker.pytorch import PyTorch
from sagemaker.inputs import TrainingInput
import boto3, sagemaker

# === 数据集路径 ===
dataset_input = TrainingInput(
    "s3://your-bucket-name/",
    distribution="FullyReplicated"
)

# === 创建 boto3 会话 ===
boto_sess = boto3.Session(
    region_name="Region"
)

# === SageMaker 会话 ===
sess = sagemaker.Session(boto_session=boto_sess)

# === PyTorch 训练任务 ===
estimator = PyTorch(
    entry_point="lerobot/train.py", 
    source_dir="/your-source-dir/", 
    dependencies=["lerobot/requirements.txt"],
    role="your-role-arn",
    instance_count=1,
    instance_type="ml.g5.2xlarge",    
    framework_version="2.1",
    py_version="py310",
    sagemaker_session=sess,
)

# === 发起训练 ===
estimator.fit({"dataset": dataset_input})

import subprocess
import sys

cmd = [
    'lerobot-train',
    '--dataset.repo_id=/opt/ml/input/data/dataset/', 
    '--dataset.video_backend=pyav',
    '--policy.type=act',
    '--output_dir=outputs/train/act_so100_test',
    '--job_name=act_so100_test',
    '--wandb.enable=false',
    '--policy.push_to_hub=false',
    '--policy.device=cuda',
    '--batch_size=2',
    '--steps=100000'
]

print(f"Running: {' '.join(cmd)}")
subprocess.run(cmd, check=True)

import boto3, os

boto_sess = boto3.Session(
    region_name="Region"  
)

bucket = ""   
prefix = ""         
local_dir = "./output"

s3 = boto_sess.client("s3")

resp = s3.list_objects_v2(Bucket=bucket, Prefix=prefix)
for obj in resp.get("Contents", []):
    key = obj["Key"]
    if key.endswith("/"): 
        continue
    local_path = os.path.join(local_dir, os.path.relpath(key, prefix))
    os.makedirs(os.path.dirname(local_path), exist_ok=True)
    s3.download_file(bucket, key, local_path)
    print(f"Downloaded: {key} -> {local_path}")

# 创建conda环境
conda create -y -n lerobot python=3.10
# 激活虚拟环境
conda activate lerobot

# step:1 Clone官方LeRobot仓库
git clone https://github.com/xiongqi123123/LeRobot-VLA.git
# step:2 安装所需依赖
conda install ffmpeg=7.1.1 -c conda-forge # 若使用conda环境
cd lerobot && pip install -e ".[feetech]"

...16-25
NUM_EPISODES = 10 <---
FPS = 30 <---
EPISODE_TIME_SEC = 60 <---
TASK_DESCRIPTION = "My task description" <---

# Create the robot configuration
camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)} <---
robot_config = SO100FollowerConfig(
    port="/dev/tty.usbmodem5AA90178121", id="my_red_robot_arm", cameras=camera_config <---
)

# Initialize the policy --31--
policy = ACTPolicy.from_pretrained("train\act_so100_testpretrained_model") <--此处修改为模型地址

# 需要修改脚本中对应模型权重目录
python RDKS_ModelRun/ACT/cpu_act_infer.py 

./pretrained_model/
├── config.json
├── model.safetensors
└── train_config.json

parser.add_argument('--act-path', type=str, default='outputs/train/act_so100_test/checkpoints/001000/pretrained_model', help='Path to LeRobot ACT Policy model.')
""" 
# example: --act-path pretrained_model
./pretrained_model/
├── config.json
├── model.safetensors
└── train_config.json
"""
parser.add_argument('--export-path', type=str, default='marcelo_test1', help='Path to save LeRobot ACT Policy model.') 
parser.add_argument('--cal-num', type=int, default=400, help='Num of images to generate')
parser.add_argument('--onnx-sim', type=bool, default=True, help='Simplify onnx or not.') 
parser.add_argument('--type', type=str, default="nash-e", help='Optional: nash-e, nash-m, nash-p, bayes-e, bayes') 
parser.add_argument('--combine-jobs', type=int, default=6, help='combie jobs for OpenExplore.')

opt = parser.parse_args([])

python RDKS_ModelRun/ACT/export_bpu_actpolicy.py \
  --dataset.repo_id=${HF_USER}/so100-test-0903 \
  --dataset.root=...... \
  --policy.type=act \
  --policy.device=cpu \
  --policy.policy=111 \
  --wandb.enable=false 

├── BPU_ACTPolicy_TransformerLayers
│   ├── BPU_ACTPolicy_TransformerLayers.onnx
│   ├── build_BPU_ACTPolicy_TransformerLayers.sh
│   ├── calibration_data_BPU_ACTPolicy_TransformerLayers
│   ├── config_BPU_ACTPolicy_TransformerLayers.yaml
├── BPU_ACTPolicy_VisionEncoder
│   ├── BPU_ACTPolicy_VisionEncoder.onnx
│   ├── build_BPU_ACTPolicy_VisionEncoder.sh
│   ├── calibration_data_BPU_ACTPolicy_VisionEncoder
│   ├── config_BPU_ACTPolicy_VisionEncoder.yaml
├── bpu_output_act_0417_2arms
│   ├── action_mean.npy
│   ├── action_mean_unnormalize.npy
│   ├── action_std.npy
│   ├── action_std_unnormalize.npy
│   ├── laptop_mean.npy
│   ├── laptop_std.npy
│   ├── phone_mean.npy
│   └── phone_std.npy
└── build_all.sh

# 下载并挂载RDKS100官方Docker工具链交付物
#CPU Docker镜像
wget -c ftp://oeftp@sdk.d-robotics.cc/oe_v3.2.0/ai_toolchain_ubuntu_22_s100_cpu_v3.2.0.tar --ftp-password=Oeftp~123$%
#GPU Docker镜像
wget -c ftp://oeftp@sdk.d-robotics.cc/oe_v3.2.0/ai_toolchain_ubuntu_22_s100_gpu_v3.2.0.tar --ftp-password=Oeftp~123$%
# 导入镜像
docker load < ai_toolchain_ubuntu_22_s100_gpu_v3.2.0.tar

# 挂载镜像执行自动编译
[sudo] docker run [--gpus all] -it -v <BPU_Work_Space>:/open_explorer REPOSITORY:TAG
# Example：
docker run -it --rm --gpus device=6 --shm-size=32g \
  -v {path/to/your/folder}:/open_explorer \
  ai_toolchain_ubuntu_22_s100_gpu:v3.2.0
  
# 运行编译
bash build_all.sh

$ tree bpu_output
.
├── BPU_ACTPolicy_TransformerLayers.hbm
├── BPU_ACTPolicy_VisionEncoder.hbm
├── action_mean.npy
├── action_mean_unnormalize.npy
├── action_std.npy
├── action_std_unnormalize.npy
├── laptop_mean.npy
├── laptop_std.npy
├── phone_mean.npy
└── phone_std.npy

python RDKS_ModelRun/ACT/bpu_act_infer.py