The hardest part of robotics isn't the hardware. It's the software that tells a robot arm how to fold a towel, a gripper how to pick up a screw, or a mobile platform how to navigate a cluttered room without a team of PhDs writing bespoke code for every task. Dream Labs, a San Francisco startup founded in 2026, is betting that a single, general-purpose model can be that brain for any robot with a camera and a motor.
In October 2026, the company emerged from stealth with a $90 million seed round and a public API for its flagship model, Dream-1. The proposition is straightforward: feed the model a live camera feed from a robot and a plain-English instruction, and it returns the motor-control commands needed to execute the task. The company is selling access to this capability starting at $3 per hour [Dream Labs, retrieved 2026].
The GR00T pedigree
The bet is backed by a team that helped build one of the industry's most watched robotics initiatives. Co-founders Joel Jang and You Liang Tan are former researchers from NVIDIA, where Jang led the world model team for Project GR00T, the chipmaker's open foundation model for generalist humanoid robots [NVIDIA Learning and Perception Research, retrieved 2026]. The broader founding team is composed of former leaders and researchers from NVIDIA's GR00T work, with some also having backgrounds in Google DeepMind's advanced robotics research [PERPLEXITY SONAR PRO BRIEF].
This pedigree gave investors confidence to write one of the largest seed checks in recent robotics history. The $90 million round was led by B Capital and Kindred Ventures, with participation from A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC, and angel investor Adrian Aoun [Forbes, October 2026]. The capital provides an unusually long runway to tackle a problem that has consumed entire research divisions for years.
The hardware-agnostic wedge
Dream Labs is not building robots. Its wedge is to become the universal control layer that sits between any robot's hardware and the work it needs to do. The Dream-1 model is designed to work across different robots, tasks, and environments, targeting bimanual robots with parallel grippers for tabletop tasks like picking up objects and folding laundry [PERPLEXITY SONAR PRO BRIEF].
The company claims the model can perform new tasks and operate in new scenes it has never seen before, a capability known as zero-shot learning. For tasks requiring more precision, users can fine-tune Dream-1 on their specific robot by uploading demonstration videos [Dream Labs, retrieved 2026]. The commercial interface is the Dream API, which exposes the model to developers and manufacturers.
Early traction appears to be on the partnership front. At launch, Dream Labs reported more than 20 hardware partners were working with it, including major names like Unitree and Robotis [Korea Post, October 2026]. This suggests a strategy of embedding its model at the manufacturer level, making it the default "brain" for new robots rolling off assembly lines.
Seed Round (Oct 2026) | 90 | M USD
The $3-an-hour economics
Pricing robot intelligence by the hour is a novel approach that aligns cost directly with utilization. The Dream API starts at $3 per hour, with users able to save up to 20% with yearly plans [Luma, retrieved 2026]. This creates a clear, scalable cost model for integrators and manufacturers, contrasting with the large upfront licensing fees or perpetual uncertainty of in-house research budgets.
For a customer like MakerMods, which sells $2,000 bimanual robot arms, the value is in solving the "last mile" problem of making purchased hardware actually productive. "Customers’ main difficulty is getting robots to perform useful work," said Ryan Chan, MakerMods co-founder and CEO [PERPLEXITY SONAR PRO BRIEF]. At $3 an hour, the calculus shifts from whether a task can be automated to whether the value of the completed work exceeds the runtime cost.
The scale problem
The technical ambition here is significant. Translating natural language and pixel data into reliable, safe motor commands across thousands of robot morphologies and millions of unpredictable environments is a frontier research problem. While the team's GR00T experience is a strong credential, moving from research demonstrations to robust, production-grade inference at scale involves a different set of challenges.
Key technical hurdles will determine if the bet pays off. Latency and reliability are non-negotiable for physical systems; a delayed or incorrect motor command can mean a dropped part or a safety incident. The model must also generalize not just to new tasks, but to the vast array of lighting conditions, camera angles, and object types found in real warehouses, labs, and homes. Finally, the fine-tuning pipeline needs to be smooth enough for non-researchers to use effectively, turning demonstrations into improved performance without introducing instability.
At scale, the failure modes are physical. A misinterpreted instruction in a logistics center could misroute packages. A confidence error in a delicate assembly task could damage components. The model's performance will be judged not just by accuracy on a benchmark, but by its mean time between failures in the field and the cost of those failures when they occur. Dream Labs' research background suggests they understand the problem space, but commercial deployment will test that understanding under real-world entropy and economic pressure.
The crowded frontier
Dream Labs is entering a field that is suddenly attracting capital and talent. The foundational model approach to robotics is being pursued by other well-funded startups and tech giants, all aiming to solve the same generalizability problem. While no direct competitors were named in the available sources, the company's success will depend on executing faster and forming deeper hardware partnerships than other entrants.
The next twelve months will be about moving from announced partnerships to integrated deployments. Evidence of paid API usage, case studies showing reduced development time for specific applications, and expansion of supported robot platforms will be the metrics to watch. With $90 million in the bank, the company has the resources to iterate quickly, but the clock is now ticking on turning research pedigree into commercial proof.
Sources
- [Dream Labs, retrieved 2026] Dream Labs | https://dreamlabs.ai/
- [Forbes, October 2026] Robot Brains For 75 Cents/Hour? Former Nvidia Researchers Raise $90 Million | https://www.forbes.com/sites/johnkoetsier/2026/10/19/robot-brains-for-75-centshour-former-nvidia-researchers-raise-90-million/
- [Korea Post, October 2026] Dream Labs, Founded by Korean Entrepreneur in His 20s, Raises $90 Million in Seed Funding | https://www.kspost.biz/en-us/articles/3061
- [Luma, retrieved 2026] Luma pricing page reference | (Source referenced in research)
- [NVIDIA Learning and Perception Research, retrieved 2026] GR00T N1 model page | https://research.nvidia.com/labs/nvidia-learning-and-perception-research/
- [PERPLEXITY SONAR PRO BRIEF] Web-grounded research brief on Dream Labs
- [RuntimeWire, October 2026] Dream Labs raises $90M to sell robot control through a $3-an-hour API | https://runtimewire.com/article/dream-labs-90m-seed-dream-1-robot-api