具身智能与机器人:DeepSeek 五層套件
让 DeepSeek 进入物理世界。这个方向资料最少 —— GitHub 上搜「机器人」大部分是聊天机器人,不是机械臂。
模型
12
工具鏈
12
參考實現
25
硬件
5
一、模型層
二、工具鏈
| 專案 | ★ | 做什麼 |
|---|---|---|
| Intelligent-Chat-Robot | 13 | 参与2025年传智杯的人工智能大模型创新应用挑战赛获得一等奖(国家级),使用Ollama客户端调用本地大语言模型(如qwen2、deepseek-r1)两种AI,集成科大讯飞API |
| Embodied-DeepSeek-Harness | 6 | Agentic framework based on DeepSeek Harness that controls your robot to excel at long hori |
| Minecraft-AIBOT | 4 | AI-Bot — 我的世界 AI 机器人伴侣 Mod ▎ ▎ 一个智能机器人伴侣 Mod(Fabric 版)。用自然语言控制机器人——聊天、跟随、战斗、挖矿、建造、合成等。 |
| pingo | 3 | pingo智能体演示平台 |
| J.A.R.V.I.S.-Holographic-Interface | 3 | An Iron Man HUD-inspired holographic interactive interface: featuring real-time camera fee |
| AI-Robot-Demo | 2 | A beginner project for LLM-based robot task planning using DeepSeek API. |
| langgraph-fastapi-chatRobot | 2 | 一个基于 LangGraph 构建的可组合对话工作流,集成 DeepSeek 模型,支持工具调用(查询天气、腾讯翻译、时间查询),提供命令行交互与 RESTful API 两种使用 |
| IAgent | 1 | A lightweight CLI tool powered by the DeepSeek API for automated Excel file manipulation a |
| L-Link-Robot-Gateway | 0 | A lightweight LLM-to-Robot instruction gateway using DeepSeek-R1 and URSim for embodied in |
| embodied-data-platform | 0 | 具身数据平台插件开发 — embodied-AI dataset toolchain (validate / quality / convert / catalog / scaff |
三、參考實現
| 專案 | ★ | 做什麼 |
|---|---|---|
| VLA-expert-skill | 49 | VLA領域知識技能:按需讀取已發布Handbook、來源分級與可證偽分析;DeepSeek每日維護來源索引,歷史信念分開標示。 |
| dsh-robotic-harness | 30 | Robotic Harness: embodied-intelligence research tools for DeepSeek Harness - robot asset i |
| jailbreak-Deepseek | 17 | A universal one-prompt jailbreak for all DeepSeek versions is theoretically described as a |
| DeepSeek-Trade-Robot | 12 | A trade robot on pumpfun use DeepSeek AI |
| YOLO-DeepSeek-Powered-Image-Manipulation | 8 | This project showcases a deep learning pipeline for image editing, integrating YOLO for ob |
| CoFlu | 8 | CoFlu is a powerful text manipulation, generation, and comparison tool. It's designed for |
| DSH-WebGIS | 7 | DeepSeek Harness (DSH) WebGIS Plugin: Map reading, map manipulation, and 3D world comprehe |
| deepseek-whalechan-robot | 3 | DeepSeek 鲸鱼娘具身机器人 🐋 An expressive Whale-chan robot in MuJoCo: walking, waving, tail motion |
| Arduino-Raspi-Car | 3 | An omnidirectional mobile robot project integrating Computer Vision (YOLOv8), Large Langua |
| ai-guide | 2 | Provide free, open access to comprehensive AI tools, guides, reviews, and resources to red |
四、硬件適配
什麼跑在什麼上。這些是能跑的真實案例,不是廠商宣傳。
| 專案 | ★ | 硬件目標 |
|---|---|---|
| awesome-llm-and-aigc | 817 | 🚀🚀🚀A collection of some awesome public projects about Large Language Model(LLM), Vision La |
| LoongForge | 816 | A high-performance framework for training LLMs, VLMs, diffusion, and embodied models on NV |
| bratan-robot | 2 | Bratan: an autonomous home robot built from a used robot vacuum, an Android phone instead |
| mimiclaw_deepseek_V4 | 2 | 基于 ESP32-S3 与 DeepSeek 的具身智能(Embodied AI)个人助手框架。支持通过飞书/Lark 实现端到端的原生大模型工具调用,可自主驱动板载 RGB、外部 |
| jieji | 0 | 解机 · 一个摄像头 + 一个麦克风 + 一个大模型,驱动一台真机器人。看→想→动,带安全层和模拟器。 |
五、踩坑
- 云端 API 做不了实时控制。机械臂控制回路要 10-100Hz,云端往返 500ms 起,只能做高层规划。
- 分层是唯一可行架构:云端大模型做任务分解,端侧小模型/规则做实时控制。
- 仿真到真实的差距(sim2real)比模型本身更难。Isaac Sim/MuJoCo 里能跑不代表实物能跑。
- VLA 模型需要动作数据,开源数据集(Open X-Embodiment 等)覆盖的机器人型号有限。
- 量化后的 VLA 在精细操作上精度掉得很明显,抓取类任务要单独验证。