rebot-mhs-lab
This server provides a safe, simulation-only MCP interface for managing a two-arm Seeed reBot hardware setup.
Discover devices — list available arms (e.g.
rebot-dm,rebot-rs), their slots, physical metadata, and safety limits.Read device state — query a device's state, joint positions, enabled status, or emergency-stop state.
Write device slots — send position updates or other slot writes; dry-run by default, with real motion only when
apply=trueandconfirmed=trueand actuators are enabled.Emergency stop — immediately latch an E-stop and disable the selected arm.
Plan two-arm handoff — generate a dry-run plan for object transfer between arms; never moves real hardware.
Enforce safety constraints — joint changes are limited (e.g. 0.15 per step), E-stop overrides motion, and all operations confirm safety before applying.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@rebot-mhs-lab List all discovered reBot devices and show their emergency stop status"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Open Hardware Agent Lab

以 Seeed reBot 双机械臂为第一套参考硬件,探索安全、可复现、模型无关的 AI-to-hardware 接口。
清乐智能 · Physical AI for Science — 访问项目网站
这是一个面向两台 Seeed Studio reBot 机械臂的安全起步项目。它先把 Anthropic 已公开介绍的 MHS 核心概念——设备发现、统一 read/write、自然语言硬件描述、设备级安全限制——做成可运行的仿真,再通过 MCP 暴露给 Claude、DeepSeek Harness 或其他智能体。
当前状态:MHS-ready prototype,不是官方 MHS 实现。 截至 2026-09-01,MHS 仍是申请制研究预览,规范与 SDK 尚未公开。拿到官方预览后,计划只替换协议/驱动适配层。
本仓库是独立社区研究项目,与 Anthropic、DeepSeek AI、Seeed Studio 或 Hugging Face 没有隶属或背书关系。
项目文档
Related MCP server: Robotics Control MCP Server
现在可以演示什么
自动发现
reBot Arm B601-DM和reBot Arm B601-RS两台虚拟设备。读取设备状态、标准化关节位置、使能状态和急停状态。
所有写操作默认 dry-run,不会移动机械臂。
运动必须同时满足:未触发急停、已显式使能、
apply=true、confirmed=true。单次关节变化限制为
0.15,越界命令在 driver 层拒绝。生成双臂物体交接计划,但当前永远只做 dry-run。
推荐学习顺序
单臂 bring-up:组装、供电、电机 ID、零点和急停。先用 MotorBridge 单独验证每个关节。
LeRobot 接口:理解
connect → get_observation → send_action → disconnect,完成 Leader/Follower 遥操作。数据闭环:录制一个简单任务、回放、检查数据,再训练 ACT/SmolVLA 等策略。
MHS 设备层:把关节状态、动作、物理属性和安全边界变成可发现的设备 slot。
MCP 编排层:让智能体读取状态、生成计划和调用受限动作;高速控制仍由 LeRobot 策略或确定性代码完成。
双臂 Demo:先做仿真交接,再做真实双臂标定、碰撞区和人工确认。
不要一开始就训练 VLA,也不要让大模型逐帧直接输出电机控制。第一阶段的目标是一个可靠、可复现、随时能急停的控制链路。
在这台 Mac 上运行仿真 Demo
cd /Users/dave/Desktop/physical_ai
uv sync --extra dev --python 3.12
uv run pytest
uv run mcp dev server.pyMCP Inspector 打开后,可以依次调用:
discover_devicesread_device(device_id="rebot-dm", slot="state")write_device(device_id="rebot-dm", slot="position", value={"shoulder_pan": 0.1})plan_two_arm_handoff(from_device="rebot-dm", to_device="rebot-rs")
第 3 步没有传入 apply=true,因此只返回预演结果。
也可以把本地 stdio server 接入支持 MCP 的宿主:
{
"mcpServers": {
"rebot-mhs-lab": {
"command": "uv",
"args": [
"run",
"--directory",
"/Users/dave/Desktop/physical_ai",
"python",
"server.py"
]
}
}
}接入 DeepSeek Harness
DeepSeek Harness 已提供官方 MCP client 插件,因此第一阶段不需要直接修改 Harness。我们把本仓库作为独立 MCP hardware server,通过 @deepseek-ai/dsh-mcp-client 接入。示例配置位于 .dsh/cordis.patch.yml。
npx @deepseek-ai/dsh webDeepSeek Harness 仍处于 developer preview,插件版本应与宿主版本保持一致。详细步骤和后续原生插件路线见 docs/DEEPSEEK_HARNESS.md。
接真实机械臂前的环境选择
当前电脑是 Apple Silicon Mac,适合运行本项目、MCP 和仿真。
B601-DM 未来可通过串口桥/MotorBridge 接入,但 Seeed 的完整 LeRobot 教程以 Ubuntu 物理机为主。
B601-RS 的 follower 教程使用 SocketCAN;真实控制建议放在 Ubuntu 22.04 或 Jetson 上,而不是直接放在 Mac 上。
推荐架构:Mac 运行 Claude/MCP 客户端,Ubuntu/Jetson 贴近机械臂运行 driver;两者通过受限的网络接口通信。
与真正 MHS 的迁移边界
目前的 SimulatedArmDriver 只提供三件事:
discover() 设备能力、物理属性、安全限制
read(slot) 状态读取
write(slot, value) 带设备级校验的写入官方 MHS SDK 可用后,新建 OfficialMHSReBotDriver 实现同样的行为,把现有 MCP 工具和双臂编排保留下来。真实 LeRobot 适配器则把标准化状态映射到 get_observation() / send_action()。
真实硬件安全门槛
固定底座、清空工作区并准备物理急停。
初次调试保持至少一米距离,限制速度、扭矩和单步变化。
禁止带电插拔电源或信号线。
断电、掉线或反馈异常后,先停止程序并重新回零,不能从旧状态继续。
真实运动不得沿用示例中的标准化位置;必须根据各自校准数据、关节限位和碰撞模型生成。
资料
授权状态
仓库目前公开用于审阅、学习和讨论,但尚未选择开源许可证。在许可证正式确定之前,请不要假定拥有复制、修改、分发或商业使用本代码的权利。这样可以在创业项目早期保留专利、双许可证和商业授权的选择空间。
Available Tools
5 toolsdiscover_devicesA
Discover the available arms, their slots, physical metadata, and safety limits.
Inputs: no named parameters. Returns: object
emergency_stopB
Latch the emergency stop immediately and disable the selected arm.
Inputs: device_id (required). Returns: object
plan_two_arm_handoffA
Build a dry-run plan for a two-arm object handoff; never moves hardware.
Inputs: to_device, from_device. Returns: object
read_deviceA
Read a device slot: state, position, enabled, or emergency_stop.
Inputs: device_id (required), slot. Returns: object
write_deviceA
Write one device slot.
Calls are dry-run by default. Applying motion requires enabled actuators plus both apply=true and confirmed=true. Emergency stop always disables motion.
Inputs: slot (required), value (required), device_id (required), apply, +1 more. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 5 tools
Each tool has a clearly distinct purpose: discovering metadata, reading state, writing state, triggering an emergency stop, and planning a two-arm handoff. There is no meaningful overlap that would cause an agent to select the wrong tool.
Most tool names follow a clear verb_noun pattern: discover_devices, read_device, write_device, plan_two_arm_handoff. The exception is emergency_stop, which is a noun-style command rather than a verb-prefixed action, making the set slightly inconsistent.
Five tools is a well-scoped size for this robotics lab server. Each tool covers a distinct operation and none feel redundant or extraneous.
The tool set covers device discovery, state reading, state writing, safety shutdown, and handoff planning. Minor gaps exist around explicitly executing a planned handoff or resetting a latched emergency stop, but these may be intentionally handled outside the MCP surface or through write_device.
Related MCP Connectors
Give the physical world an API: cameras become typed endpoints your agent can read and act on.
Connect your AI to your own devices, 3D printers, factory tools and a shared agent task pool.
Read-only finance and operations controls for AI agents with evidence and safe next actions.
Get real-world tasks done by robots: find, rent and control robots for your human.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables AI assistants to control Universal Robots through real-time connection management, status monitoring, and precise joint/linear motion control. Provides safe robot operation with built-in collision detection and simulation mode for development without physical hardware.6GPL 3.0
- AlicenseNot gradedqualityCmaintenanceEnables AI agents to control hardware devices like Arduino, Raspberry Pi, 3D printers, CNC machines, and custom robots via serial ports and HTTP. Provides tools for device discovery, command sending, sensor reading, servo control, G-code execution, and emergency stops with safety features.37 PyPIMIT
- AlicenseAqualityBmaintenanceEnables LLM agents to safely inspect, plan, preview, and execute motions on ROS 2 manipulators, with server-enforced safety envelopes covering joint limits, workspace constraints, force/torque thresholds, rate limits, e-stop, human approval, and audit logging.2247 PyPIApache 2.0
- AlicenseAqualityCmaintenanceEnables AI assistants in MCP clients to read state and move a calibrated SO-101/SO-ARM101 robot arm through LeRobot, with per-joint safety limits, named poses, joint moves, and gripper controls.82MIT