Dream Labs raises $90M seed for a robot-brain API from ex-NVIDIA GR00T researchers
Dream Labs, a San Francisco startup founded in 2026 by former leads of NVIDIA's GR00T robotics research, raised $90 million in seed funding led by B Capital alongside Kindred Ventures, the company said in a Business Wire release on Oct 19.
It also opened Dream API, which turns a robot's camera feed and a plain-English task into motor commands for arms costing about $2,000 to $10,000. It is the third-largest robotics seed round of $20 million or more in MEKAWIRE's data since October 2025.
Dream Labs says a robot arm it has never seen can do useful work within minutes, and it has just been given $90 million to prove it. The San Francisco startup, founded in 2026 by former leads of NVIDIA's GR00T robotics research, said in a press release distributed by Business Wire on Oct 19 that it raised a seed round led by B Capital alongside Kindred Ventures, with A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC and Adrian Aoun among the other investors. It did not disclose a valuation. [Company profile](/companies/dream-labs).
The money buys almost no hardware. The company said it will spend the round primarily on compute, to widen the tasks and robots its model can handle and to serve more users. The question this article answers is whether a robot brain sold through an API can become the default for cheap arms before far larger rivals, or NVIDIA's own open models, get there first.
The product is an API, not a robot
Alongside the financing, Dream Labs opened three products. Dream API takes a robot's camera images, its current joint state and a plain-English task, and returns the actions the robot should execute; the company said its models run in data centres around the world so robots respond quickly wherever they are. Fine-tuning as a Service lets a robot maker submit a small batch of demonstrations and receive a tailored model within hours; the company's website says a few hundred demonstrations adapt the model to a new body. Robot Playground lets anyone with a browser type commands to a robot in the company's San Francisco office and watch it respond live, unedited.
The flagship model, Dream-1, today handles tabletop work such as picking and placing objects and folding laundry, according to the release. It needs no extra data to drive a two-armed robot with parallel grippers, and the company refines it for each of its more than 20 hardware partners. The robots it supports cost about $2,000 to $10,000. Support for wheeled mobile robots and dexterous hands is to follow, with no date given. Ryan Chan, chief executive of MakerMods, which sells $2,000 bimanual arms, said in the release that his robot was running within minutes and attempting tasks it had never been programmed for.
None of that comes with a benchmark figure, which makes the mechanism worth unpacking.
How a world action model differs from a VLA
Dream Labs' team page says the group released DreamZero in February 2026, the paper that introduced what it calls World Action Models, and Dream-1 about six months later. The release describes the method plainly: Dream-1 is trained on video of people doing everyday tasks, predicts what will happen next in the scene, and derives the robot's action from that prediction. A conventional vision-language-action model ([MEKAWIRE explainer](/news/what-is-a-vla-model-robot-foundation-models-explained-rt-2-pi0-gr00t-gemini-robotics-2-94bn-raised-2026-20261017)) maps images and text straight to motor commands and is usually trained on logs from specific machines. Dream Labs argues that because its training signal is human video rather than one robot's data, skills carry across hardware. MEKAWIRE's [world model explainer](/news/what-is-a-world-model-in-robotics-explained-amd-world-labs-cosmos-praxis-1-44-papers-20261019) covers the wider family this belongs to.
The release gives no task-success rates, no latency figures and no API prices. Running inference in remote data centres also means every action depends on a network link, a trade-off factory buyers will weigh.
Third-largest robotics seed round in MEKAWIRE's data
- Walden Robotics (US)
- Jul 15, 2026
- $300M
- Deviation Capital
- General Intuition (US)
- Oct 16, 2025
- $133.7M
- General Catalyst
- Dream Labs (US)
- Oct 19, 2026
- $90M
- B Capital
- Generative Bionics (IT)
- Dec 9, 2025
- $81M
- CDP Venture Capital
Verified rows labelled seed, $20M and above. Dream Labs / General Intuition = 90 / 133.7 = 67%; / Walden = 90 / 300 = 30%; / Generative Bionics = 90 / 81 = 1.11x. Disclosed, not estimated.
As of Oct 20, 2026
Among rounds labelled seed of at least $20 million since October 2025 in MEKAWIRE's verified data, Dream Labs ranks third. It equals 67% of General Intuition's seed and 30% of Walden Robotics', and is 1.11 times Generative Bionics'.
- Dream Labs
- Seed
- Oct 19, 2026
- $90M
- Not disclosed
- General Intuition
- Seed
- Oct 16, 2025
- $133.7M
- Not disclosed
- Physical Intelligence
- Series B
- Nov 20, 2025
- $600M
- $5.6B
- Skild AI
- Series C
- Jan 14, 2026
- $1.4B
- $14B
- Generalist AI
- Growth
- Jun 4, 2026
- $400M
- Not disclosed
MEKAWIRE verified rows. Skild AI Series C / Dream Labs seed = 1,400 / 90 = 15.6x. Valuations only where disclosed.
As of Oct 20, 2026
Against the best-funded robot-brain companies it is a small cheque: Skild AI's January 2026 Series C alone was 15.6 times Dream Labs' seed. B Capital led Apptronik's $350 million Series A in February 2025 and its $53 million additional close, and joined the $520 million extension in February 2026. SV Angel was in Skild AI's $300 million Series A in July 2024, and Kindred Ventures in Nuro's $203 million Series E in August 2025.
MEKAWIRE analysis
Dream Labs is selling robot intelligence as a metered cloud service for the cheapest arms on the market; its edge is reach into $2,000 hardware, not model scale, and its risk is that the model layer is being given away.
The arithmetic shows where the bet sits. The $90 million seed equals the list price of 9,000 robots at the top of the supported range, or 45,000 at the bottom ($90 million divided by $10,000 and by $2,000). The company is not buying a fleet; it is buying compute and courting hobbyists, labs and small makers through a browser playground and fast fine-tuning. Those are distribution moves aimed at buyers who cannot afford months of specialist integration.
The strongest counter-argument comes from the founders' former employer. NVIDIA releases its GR00T models openly and earns its return on data-centre GPUs ([MEKAWIRE explainer](/news/nvidia-robotics-strategy-explained-2026-jetson-thor-gr00t-cosmos-free-models-sell-data-center-gpus-20-robot-startups-backed-20261017)). If an open model is good enough on a $2,000 arm, a paid API has to win on support, fine-tuning speed and reliability rather than on access. Physical Intelligence and Skild AI also have many times more capital.
Bull case: partner makers ship Dream-1 as the default brain, usage grows with every cheap arm sold, and the human-video approach keeps improving without a fleet of data-collection robots. Bear case: open models close the gap, factories refuse cloud-dependent control, and the API stays a developer tool with little revenue.
What would change our view: published success rates on partner hardware, a public price list, and a working release for mobile robots or dexterous hands. The next dated marker is NVIDIA's GTC Berlin keynote on Oct 21, which may show how far NVIDIA's own open robot models have moved. The open question is whether the first robot makers to bundle Dream-1 will pay for it or expect it free.
- B Capital, alongside Kindred Ventures
- $90 million seed, valuation not disclosed
- Dream-1 model via Dream API, fine-tuning as a service, browser Robot Playground
- Former leads of NVIDIA's GR00T research; co-founder and CEO Joel Jang
- More than 20 partners; supported robots cost about $2,000 to $10,000
- Primarily compute (company statement)
- Third-largest seed round of $20M+ in MEKAWIRE robotics data since Oct 2025
- A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC, Adrian Aoun
Why it matters
Dream Labs is one of the first robot foundation-model companies to sell access through a public self-service API rather than through enterprise pilots or bundled hardware. If it works, the cheapest robot arms gain a general-purpose controller without months of integration, and model companies compete on developer adoption the way cloud AI providers do.
The round is also a talent signal: the founders led GR00T research at NVIDIA, the company whose open robot models and Jetson hardware set the reference stack for most humanoid and manipulation start-ups in MEKAWIRE's data.
Rival analysis
Physical Intelligence ($600 million Series B at $5.6 billion, November 2025) and Skild AI ($1.4 billion Series C at $14 billion, January 2026) build general robot models with far more capital and sell mostly to partners and enterprises. Generalist AI ($400 million, June 2026) and General Intuition (seed $133.7 million, October 2025) bet on large proprietary data sets. Dream Labs differs in distribution: an open API, a browser playground and hardware priced at $2,000 to $10,000.
NVIDIA is both the founders' former employer and the most direct substitute, because it releases GR00T models openly. Galaxea, a Beijing maker whose logo appears among Dream Labs' early partners, also builds its own VLA foundation models.
Valuation context
No valuation was disclosed, so MEKAWIRE records none. The $90 million seed is the third-largest seed round of at least $20 million in MEKAWIRE's robotics data since October 2025, after Walden Robotics ($300 million) and General Intuition ($133.7 million).
The seed is 6.4% of Skild AI's January 2026 round and 15% of Physical Intelligence's November 2025 Series B by amount (90/1,400 and 90/600). Any later priced round will be read against those two disclosed post-money valuations.
Supply-chain implications
The cost base is compute, not components: the company said most of the money goes to compute and that its models run in data centres around the world. That ties gross margin to cloud GPU prices and, indirectly, to NVIDIA.
On the hardware side, the supported robots are low-cost arms from more than 20 partners, many of them Chinese makers of bimanual and wheeled platforms. Dependence on partner hardware quality and on network latency are the two operational exposures.
Signals to watch
Oct 21, 2026: NVIDIA GTC Berlin keynote, for any update to the open GR00T models that compete with Dream-1.
No date given: Dream Labs' release of wheeled mobile robot and dexterous hand support, and the first published task-success figures on partner robots.
No date given: a public price list for Dream API and fine-tuning, which will show whether the business is usage-based or licensed through makers.
Analyst view
Thesis: Dream Labs is a distribution bet on cheap arms. Confidence: medium. The team's research record and the investor list are strong, but the release contains no performance numbers, prices or customer counts, and open models from NVIDIA set a low price ceiling.
We would raise confidence on published success rates across several partner robots and on evidence that makers pay for the API rather than receive it free.
Questions you should be asking
What is Dream-1's task-success rate on partner hardware it has not been fine-tuned for?
How is Dream API priced, and who pays: the robot maker or the end user?
What latency does cloud inference add per action, and is an on-device version planned?
How does Dream-1 relate to NVIDIA's GR00T models in licensing and training data?