12/30/25

25115 Youibot 的 One-Brain-Multi-Form 架構

 Youibot 的 One-Brain-Multi-Form 架構是一種具身智能系統,使用單一通用 AI 大腦(MAIC 模型)驅動多種機器人形態,實現跨場景的高泛化操作。facebook+2

核心組件

架構以 MAIC(Mobile AI Comprehension)為「一腦」核心,基於 VLM(視覺語言模型)與 MoE(混合專家模型)訓練於多模態數據集與「Data Ocean」,具備場景感知、語言理解、任務規劃與決策能力。youtubeyouibot

  • 大腦層:負責高層邏輯,如環境理解、任務分解與路徑優化,支持多模態輸入(視覺、語言、感測器數據)。36kr

  • 小腦層:處理低層動作執行、錯誤修正與即時反饋,確保精準運動控制。36kr

  • 多形態層:支援輪式人形(Vortex)、雙足、四足、履帶等形態,單腦適配不同硬體,無需重訓。youibot+1

運作流程

  1. 感知與理解:多感測器(LiDAR、相機)輸入數據至 MAIC,模型透過 VLM/MoE 解析場景、物件與指令(如自然語言任務)。youibotyoutube

  2. 規劃與決策:大腦生成任務序列與路徑,支援多機器人協作(YOUIFleet 調度 200+ 台),自動選擇最適形態(如輪式巡檢、臂式操作)。linkedin+1

  3. 執行與學習:小腦驅動硬體動作,邊緣計算即時調整;全流程閉環學習,從數據收集到持續優化,實現端到端擴展。www-web.itigeryoutube

  4. 適應與泛化:單腦跨形態轉移(如從倉儲輪式到半導體人形),降低部署成本,支持半導體、能源等工業任務。globaltimes+1

優勢與應用

此架構實現「一腦多機、多能」,單機器人高精度操作,多形態無縫切換,提升穩定性與效率,已落地智能製造、巡檢與物流。youibot+1

  1. https://www.facebook.com/ChinaFocus2015/posts/a-humanoid-robot-which-is-reported-to-be-the-worlds1st-cross-scenario-cluster-op/1080586040946310/
  2. https://zh.youibot.com/news/183.html
  3. https://www-web.itiger.com/news/1149621111
  4. https://www.youtube.com/watch?v=6TDTa1bxEUg
  5. https://eu.36kr.com/en/p/3013366877496832
  6. https://en.youibot.com/news/youibots-vortex-wheeled-humanoid-robot-nominated-for-2025-iera-award.html
  7. https://www.globaltimes.cn/page/202508/1340237.shtml
  8. https://tw.linkedin.com/company/youibot
  9. https://www.anews.com.tr/tech/2025/08/06/china-unveils-1st-humanoid-robot-capable-of-operating-across-scenarios
  10. https://claire-chang.com/category/ai/
  11. https://botpress.com/tw/blog/multi-agent-systems
  12. https://www.uso.com.tw/portal_b1_page.php?owner_num=b1_55912&button_num=b1&cnt_id=19104
  13. http://jqr365.youfuyun.com/mobile/index.php?moduleid=14&itemid=47&r=8528
  14. https://www.ithome.com.tw/news/171379
  15. https://www.tairoa.org.tw/uploadfiles/file/journal/AIR54%E6%9C%9F-%20%E6%95%B8%E4%BD%8D%E8%99%9F%20(2025.09%E7%99%BC%E8%A1%8C).pdf
  16. https://info.taiwantrade.com/biznews/%E5%85%B7%E8%BA%AB%E6%99%BA%E8%83%BD%E6%A9%9F%E5%99%A8%E4%BA%BA%E5%8A%A0%E9%80%9F%E7%99%BC%E5%B1%95-2768312.html
  17. https://www.youtube.com/watch?v=Psc4mBqjk-w
  18. https://cics.aeust.edu.tw/var/file/16/1016/img/327607107.pdf
  19. https://www.youtube.com/watch?v=U4AUOQS4Bw8
  20. https://www.aa.com.tr/en/asia-pacific/china-unveils-1st-humanoid-robot-capable-of-operating-across-scenarios/3652111

Youibot One-Brain-Multi-Form 架構中的 MAIC 模型(視覺語言模型 VLM 核心)訓練流程分為預訓練、混合專家微調與具身強化學習三階段,依賴自建「Data Ocean」高品質多模態工業數據集,實現跨形態泛化。xinhua+2

預訓練階段

從預訓練 VLM(如基於 Transformer 的多模態骨幹)初始化,輸入工業場景數據(圖像、點雲、語言指令、動作軌跡)。

  • 使用對比學習(CLIP-like)對齊視覺語義與語言,融合 LiDAR/相機數據訓練環境感知。hkexnews+1

  • 數據規模:數十萬小時真實工廠視頻與感測器記錄,涵蓋半導體潔淨室、汽車組裝線等,目標為語義定位與物件辨識。ofweek+1

混合專家微調(MoE)

引入 MoE(Mixture-of-Experts)架構,針對工業任務專化。

  • 多模態融合模塊「靈犀」訓練:整合圖像、點雲與自然語言,學習「視覺語義感知-行為決策」映射,支持少樣本遷移。finance.sina+1

  • 監督式微調(SFT):高品質指令-示範對(如「運輸零件至站點 A」),優化元任務泛化,每 epoch 迭代數千任務序列。hkexnews

具身強化學習階段

端到端閉環訓練「決策-關節控制」,模擬與實機並行。

  • 強化學習(RL)與模仿學習:從專家示範生成動作,獎勵物理穩定性與任務成功率;模擬器(Gazebo/ROS2)加速,實機驗證 ±2mm 精度。finance.sina+1

  • 持續學習:部署後收集互動數據回饋「Data Ocean」,閉環優化,支持多形態轉移(如輪式到人形)。xinhua+1

此流程確保 MAIC 在複雜工業環境的毫秒級推理,已落地晶圓廠批量應用。ofweek+1

  1. http://www.bj.xinhua.org/20250411/55a88dbff2db4449848b511c318aae9c/c.html
  2. https://finance.sina.com.cn/roll/2025-04-11/doc-inesunpz2642112.shtml?froms=ggmp
  3. https://www1.hkexnews.hk/app/sehk/2025/107714/a125902/sehk25092602594_c.pdf
  4. https://m.ofweek.com/ai/2025-06/ART-201712-8500-30665082.html
  5. https://arxiv.org/html/2503.09871v1
  6. https://rohitbandaru.github.io/blog/Foundation-Models-for-Robotics-VLA/
  7. https://learnopencv.com/vision-language-action-models-lerobot-policy/
  8. https://www1.hkexnews.hk/app/sehk/2025/107714/documents/sehk25092602577.pdf
  9. https://aman.ai/primers/ai/VLM/
  10. https://arxiv.org/html/2505.23450v1
  11. https://www.youtube.com/watch?v=mpuRca2UZtI
  12. https://www.marvik.ai/blog/from-words-to-actions-the-rise-of-vision-language-action-models-in-robotics
  13. https://kh.linkedin.com/company/youibot
  14. https://www.roboticsproceedings.org/rss21/p076.pdf
  15. https://www.sciencedirect.com/science/article/abs/pii/S1566253525007249
  16. https://www.leaderobot.com/news/5674
  17. https://www.sciencedirect.com/science/article/pii/S2095809925003820
  18. https://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-68.pdf
  19. https://m.36kr.com/p/3221051625137411
  20. https://www.youtube.com/watch?v=AVialKJVDfE
  21. https://github.com/jonyzhang2023/awesome-embodied-vla-va-vln
  22. https://www.itopmarketing.com/info20686
  23. https://ee.kaist.ac.kr/en/research-achieve/ee-prof-changdong-yoos-research-team-develops-next-generation-reinforcement-learning-frameworks-for-physical-ai-erl-vlm-and-plare/
  24. https://arxiv.org/html/2508.13073v1
  25. https://mit-realm.github.io/VLM-gcbfplus-website/files/Online_planning_using_VLMs.pdf
  26. https://openaccess.thecvf.com/content/CVPR2025/papers/Li_Robotic_Visual_Instruction_CVPR_2025_paper.pdf
  27. http://www.ciche.org.tw/wordpress/wp-content/uploads/2025/12/%E5%85%A8%E6%9C%AC-%E5%9C%9F%E6%9C%A8%E6%B0%B4%E5%88%A9-%E7%AC%AC52%E5%8D%B7%E7%AC%AC5%E6%9C%9F-%E9%9B%BB%E5%AD%90%E6%9B%B8.pdf
  28. https://claire-chang.com/category/ai/
  29. https://docs.oneai.twcc.ai/p/xKNcU3O5D?print-pdf
  30. https://www.tairoa.org.tw/uploadfiles/file/journal/AIR54%E6%9C%9F-%20%E6%95%B8%E4%BD%8D%E8%99%9F%20(2025.09%E7%99%BC%E8%A1%8C).pdf
  31. https://www.gsma.com/about-us/regions/greater-china/5g-in-sector-members-and-participants/
  32. https://papers.nips.cc/paper_files/paper/2024/file/8e6f3d53b2bef98fce17e699557f5f11-Paper-Conference.pdf
  33. https://eu.36kr.com/en/p/3477350780541316
  34. https://arxiv.org/html/2505.03181v1