Dream Labs
Category: Robotics / Embodied AI
Robot foundation model startup building Dream-1, a World Action Model that controls robots it has never seen, available through a public API. Dream Labs was founded in 2026. The company is led by Joel Jang. Based in San Francisco, United States. Total funding raised: $90M. Latest round: Seed. Key investors include B Capital, Kindred Ventures, A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC, Base Ventures, Thomas Wolf, Adrian Aoun.
- Founded
- 2026
- Headquarters
- San Francisco, United States
- Total funding
- $90M
Value proposition
A general-purpose robot brain that works zero-shot across hardware from different manufacturers, so a new robot can do useful tasks within minutes. No fine-tuning, no robot-specific demonstrations and no ML expertise are needed. Pricing starts at $3/hour via API.
Products and solutions
Dream-1 (World Action Model for robot control), Dream API (public API for robot foundation model inference), Robot Playground (browser-based remote control of robots in Dream Labs' SF office).
Unique value
Billed as the first publicly available API for a robot foundation model. It offers zero-shot embodiment generalization to unseen robots, tasks and environments, at low cost from $3/hour.
Target customer
Robot hardware makers, robotics researchers and engineers, developers, prosumers and small teams using low-cost ($2K-$10K) robots, especially bimanual arms with parallel grippers.
Industries served
Robotics, consumer/home robotics, research labs, manufacturing and pick-and-place tabletop manipulation
Technology advantage
Founders created the World Action Model (WAM) paradigm with DreamZero (Feb 2026, 14B-parameter WAM on a video diffusion backbone, more than 2x the generalization of SOTA VLAs). The model is trained mainly on human everyday-task video rather than robot-specific data, which lets capabilities transfer across hardware. The team behind NVIDIA GR00T N1, DreamGen and DreamDojo. Dream-1 reportedly ranked top in RoboCurve independent evaluation.
How they differentiate
Uses a World Action Model (predicts world dynamics plus actions from video) instead of a vision-language-action model. Offers a public, low-cost pay-per-use API with zero-shot transfer to unseen robots. Focuses on cheap consumer/prosumer robots rather than enterprise or proprietary humanoids.
Main competitors
Physical Intelligence, Skild AI, Google DeepMind Gemini Robotics
Key partnerships
20+ robot hardware partners including Unitree and Robotis, beta-testing partners include Hugging Face, KAIST, Scale AI, AgiBot, RoboCurve, Anvil Robotics, BitRobot, Mecka, XDOF
Notable customers
MakerMods
Major milestones
2026-02: Founders released DreamZero (introduced World Action Models) at NVIDIA, 2026-04: Company incorporated, 2026-10: Emerged from stealth with $90M seed led by B Capital, launched Dream-1, Dream API and Robot Playground.
Growth metrics
Supports 20+ hardware partners; beta-tested across 12 external sites; Dream-1 performs tasks zero-shot on unseen bimanual parallel-gripper robots.
Market positioning
Developer-facing, open-access robot brain API for the low-cost robot hardware wave. Positioned as the 'operating system' intelligence layer for everyday robots.
Geographic focus
Global (US-headquartered; partners and beta testers across US, China, Korea, Southeast Asia, Europe)
Patents and IP
Research publications: DreamZero (World Action Models), DreamDojo, DreamGen, LAPA (authored by founders while at NVIDIA GEAR Lab)
About Joel Jang
Ex-NVIDIA Senior Research Scientist, GEAR Lab (led world model team for Project GR00T; led DreamZero, DreamDojo, DreamGen, LAPA research); Korea University BS Computer Science; KAIST MS Artificial Intelligence. Coined the term 'World Action Model'.
Latest news about Dream Labs
More Robotics / Embodied AI companies
Official website: https://dreamlabs.ai