Dream-1 / Dream API

Robot foundation model from Dream Labs, a San Francisco startup founded by former NVIDIA GR00T researchers (CEO Joel Jang) that came out of stealth on October 19, 2026 with a $90M seed co-led by B Capital and Kindred Ventures. Dream-1 is a World Action Model trained on video of people doing everyday tasks: it takes three camera views (scene + both wrists), a robot state, and a plain-language instruction, and returns end-effector poses plus gripper state that each robot's own inverse-kinematics solver converts to joint motion — so skills transfer across hardware without per-robot retraining. Dream API is public: Dream-1 at $3/hour, Dream-1-flash at $2/hour, with Fine-tuning-as-a-Service and a browser-based Robot Playground driving a live robot in Dream Labs' SF office. Current scope is deliberately narrow: tabletop pick/place and folding on bimanual robots with parallel grippers, zero-shot on robots it was never trained on (company reports 27 unseen morphologies tested, including a 2008 PR2); wheeled bases and dexterous hands are roadmap. 20+ hardware partners reported. Verification level: limited — capability figures are maker-reported and worth confirming directly. Figures here trace to the maker's site (dreamlabs.ai); confirm current specs there before purchase. This entry sits in the Open-Source Projects segment of the index.

VERIFIED 2026-10-27 via company announcement (X @dreamlabs_ai 2026-10-19) + Ventureburn/42how/Wowtale corroboration: $90M seed (B Capital + Kindred Ventures, with A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC, Adrian Aoun); API pricing $3/hr Dream-1, $2/hr flash; 20+ hardware partners; 27 unseen robot morphologies tested; independent eval by robocurve cited by company (RoboDojo-RC Tier 1 59 vs 48). Pricing discrepancy in some coverage ($0.75/hr lite) — used company-consistent $3/$2.