Fo_

A foundation model for forklifts, trained on real work. 從真實作業中學習的堆高機基礎模型。

Fo learns from teleoperated forklifts working in real warehouses — synchronized video, controls, and human judgment, recorded every operating hour. Fo 向在真實倉庫工作的遠端遙控堆高機學習——影像、操作與人的判斷同步記錄,每運行一小時持續累積。

Fo-0.1 · simulation stage模擬階段

Why forklifts

Physical AI is starved of real-world data. Forklifts produce it every day. 實體 AI 缺的是真實世界的資料。堆高機每天都在產生它。

Foundation models for the physical world fail on data: robot demos are scripted, lab hours are scarce, and simulation alone doesn't transfer. Forklifts are different — they move real loads, in real warehouses, under real constraints, every single day. And where a forklift is teleoperated, every hour of work is also an hour of labeled demonstration by a skilled human operator. 物理世界的基礎模型敗在資料:機器人展示是照稿演出、實驗室時數稀少、僅靠模擬無法遷移到現實。堆高機不同——它每天在真實倉庫、真實限制下搬運真實貨物。而在遠端遙控的堆高機上,每一小時的作業,同時就是熟練操作員一小時的帶標註示範。

The flywheel

The service funds the model. The model improves the service. 服務養模型,模型強化服務。

RUF, our teleoperation service, earns revenue at working warehouses while collecting training data. Fo learns from it — and as Fo takes over more of the driving, one operator covers more sites and the service gets cheaper. 我們的遠端遙控服務「RUF」在營運中的倉庫創造營收,同時累積訓練資料。Fo 從中學習——Fo 分擔的駕駛越多,每位操作員能負責的現場就越多,服務也隨之更便宜。

01 · RUF
Operate運行
Teleoperated forklifts work paid shifts in real warehouses遠端堆高機於真實倉庫有償運行
02 · Collect
Record記錄
Multi-camera video, controls, and interventions — synchronized影像、操作與介入同步記錄
03 · Label
Curate整理
Skilled operators' judgment becomes training data熟練判斷成為訓練資料
04 · Fo
Train訓練
The foundation model learns forklift work基礎模型學會堆高作業
05 · Scale
Automate自動化
One operator, more sites — service cost drops一人顧更多現場,成本下降
cheaper service → more sites → more data → better model更便宜 → 更多現場 → 更多資料 → 更好的模型

Fo-0.1 · today

Where Fo stands now — honestly. Fo 的現況——誠實地說。

Fo-0.1 · Isaac Sim

VERSION
Fo-0.1
STAGE
Simulation. Fo-0.1 inserts forks into pallets in sim.模擬階段。可在模擬環境中將貨叉插入托盤。
DATA SOURCE
Teleoperation logs from live warehouse sites in Japan (via RUF), plus simulation日本實際倉庫的遠端操作紀錄(透過 RUF)+模擬
PHILOSOPHY
We version openly: 0.1 → 0.2 → … No demos we can't reproduce.公開刻畫版本:0.1 → 0.2 → … 不展示無法重現的內容。

Roadmap

Versioned like a model, because it is one. 作為模型,以版本迭代成長。

Fo-0.1 — now現在

Pallet engagement in simulation. Data pipeline connected to live RUF sites.模擬中的托盤對位插取。資料管線已連接運行中的 RUF 現場。

Fo-0.2 —

Learning from accumulated real-operation data across sites and seasons.從跨現場、跨季節累積的實際營運資料中學習。

Fo-0.x —

Hybrid operation: autonomous driving with human remote intervention. One operator, many sites.邁向混合營運:自動行駛×人為遠端介入。一位操作員,照看多個現場。

No dates on purpose — versions ship when they reproduce.刻意不寫日期——版本在可重現時發布。

Build Fo with us. 與我們一起打造 Fo。

We're looking for research engineers (robot learning / VLA / Sim2Real) and R&D partners — forklift OEMs, warehouse operators, and universities — who want a foundation model grounded in real work. 我們正在尋找研究工程師(機器人學習/VLA/Sim2Real)與研發夥伴——堆高機製造商、倉儲營運商與大學——一起打造根植於真實作業的基礎模型。

Contact us聯絡我們 About RUF →關於 RUF →