Dream Labs Launches Dream-1 Robot Model and Public Dream API
Dream Labs launched Dream-1, a robot foundation model trained on human videos, and made it available globally through the public Dream API. The company says Dream-1 can control compatible bimanual robots without robot-specific fine-tuning or demonstrations. Dream Labs reported zero-shot performance on 27 unseen robots in simulation and four real robots, though the supplied reports do not independently verify the results. Dream Labs also raised a $90 million seed round led by B Capital and Kindred Ventures. The main Dream API model was launched at $3 per hour.
Dream Labs calls Dream-1 an important step toward general-purpose robot intelligence, arguing that zero-shot transfer may let compatible robots begin with useful capabilities without collecting robot-specific data. The supplied reports do not independently verify the company’s performance or generalization claims.
機器人AI新創公司Dream Labs推出Dream-1模型,官方提到,即使沒有新機型的訓練資料,也能控制不同的雙臂機器人完成抓取、堆疊及擺放物品等任務。該模型已開放API供開發者測試,目前主要支援配備平行夾爪的雙臂機器人。
不同機器人的手臂長度、關節位置及夾爪設計各不相同,因此用某款機器人訓練的模型,換到其他機型時,通常未必能直接使用。Dream-1採用世界動作模型(World Action Model)技術,透過機器人頭部及兩側手腕的攝影機畫面,配合文字指令預測接下來的影像變化與動作。模型先預測夾爪的目標位置、朝向及開合狀態,再由機器人本身的控制軟體將指令轉換為各關節的動作,減少針對不同機型重新訓練模型的需求。
Dream Labs公布的測試涵蓋27款未參與訓練的模擬機器人,以及4款實體機器人。模擬測試中,每款機器人進行300次任務測試,一半檢查能否按照文字指令完成任務,另一半測試沒有學過的新任務。不過,Dream-1雖不必針對新機型重新訓練,就能執行指定任務,但實際表現仍因機型而異。
#Dream Labs#Dream-1#Dream API#B Capital#Kindred Ventures#fal