Dream Labs Dream-1 robot model: $90M seed, API from $3 an hour, claims unverified

Dream Labs, a San Francisco robotics startup, said on 19 October 2026 that it raised a $90 million seed round and opened a public API for Dream-1, a model that turns camera images and a plain-English instruction into movements for two-armed robots. The company calls Dream-1 a "World Action Model": it predicts how a scene will change and what the robot should do, together. For developers it matters because it is sold by the hour over the internet, so a small team with a cheap robot arm can try it without training its own model. Most of the performance claims, though, come from the company itself.

Key points

• Dream Labs announced a $90 million seed round on 19 October 2026, led by B Capital and Kindred Ventures, together with a public API for its Dream-1 model.
• Dream-1 predicts video and robot actions together; it takes three camera views (a head view and a wrist view on each arm) and outputs end-effector poses or joint angles, according to Dream Labs.
• Dream Labs lists Dream-1 at $3 per hour; its lighter model is listed at $2 per hour on the company's page, while two news reports give 75 cents per hour (see below).
• The company says an independent evaluation by Robocurve, dated 18 October 2026, scored Dream-1 above the next best models; the page calls the model a research beta that "has not been validated for safety-critical use."
• It currently handles tabletop tasks such as picking, placing and folding laundry on bimanual robots with parallel grippers.

What happened

On 19 October Dream Labs announced the seed round and, the same day, launched the Dream API for robots. SiliconANGLE reports that B Capital and Kindred Ventures jointly led the round, with A.Capital, Unusual Ventures, Wing VC, SV Angel, Scribble VC and the angel investor Adrian Aoun also taking part. Joel Jang is co-founder and chief executive; TechFundingNews reports that he previously was a senior research scientist at NVIDIA's GEAR Lab, where he led the world-model team for Project GR00T. The funding will mainly pay for compute, according to the same reports.

The research page for Dream-1, dated October 2026, describes the product. The model works with a bimanual robot with parallel grippers. Developers connect the robot, use a Python client, select the dreamlabs/dream-1 checkpoint, send in observations and receive actions. Each call returns 1.6 seconds of actions, which the company prices at about $0.0013 per action for Dream-1 and $0.0009 for the lighter Dream-1-flash. A browser-based "Robot Playground" lets people type commands and watch a live robot respond, and a fine-tuning service takes demonstration videos from customers who need a skill the base model lacks.

SiliconANGLE says the company works with more than 20 hardware makers whose machines cost between $2,000 and $10,000. Dream Labs' stated next steps are wheeled mobile robots and fully dexterous, human-like hands.

Where the sources disagree on price

Dream Labs' own page lists Dream-1 at $3 per hour and Dream-1-flash at "$2 per hour". SiliconANGLE and TechFundingNews (which attributes the figure to Forbes) instead report a lightweight version at 75 cents per hour. We could not tell from the sources whether the 75-cent figure describes a different tier, an earlier price or an error, so check the current price list on the company's platform before planning around it.

Background

The idea of a world action model comes from the team's earlier DreamZero work, published in February 2026 and linked from the Dream-1 page. According to TechFundingNews, that paper described a 14-billion-parameter world action model that adapted to a new robot with 30 minutes of play data. We could not open the DreamZero paper's page ourselves, so we treat these details as reported, not confirmed. The Dream-1 page does not state its own parameter count.

The conventional route to teaching a robot a task is to collect demonstrations on that specific robot, often by a human operating it, and train a policy from them. The cost of that data is the bottleneck. Dream Labs' bet, as SiliconANGLE describes it, is to treat robot skills as a video prediction problem and learn from "human egocentric training data": video of people stacking shelves or packing boxes, recorded from their own point of view. Because people's hands and a robot's grippers do similar things, the model predicts where each gripper should be and whether it should be open or closed, and leaves the joint-level motion to a solver for each robot.

SiliconANGLE estimates that more than 5 million industrial robots operate in factories and warehouses worldwide, most built for one task and hard to retrain. TechFundingNews cites the International Federation of Robotics counting a record five million industrial robots operating in factories in 2025. Cheap dexterous arms, by contrast, now cost a few thousand dollars, so the software is the missing piece. In Jang's words, "The hardware is ready, but what's been missing is a brain any robot can use."

How it works

According to the company's page, Dream-1 takes input from three cameras and can output either end-effector poses or joint angles. Predicting the end-effector pose, the target position and orientation of each gripper, rather than joint commands is the design choice that lets one model serve different arms: joint commands that work on one arm will not necessarily work on another. A per-robot inverse-kinematics (IK) solver converts the pose targets into joint movements, and the page states that performance depends on each robot's IK solver. A getting-started tutorial and an IK solver guide are linked from the page.

The page lists these limits in its own words or close to them:

• It controls where the gripper ends up, not how the arm gets there, which can matter near obstacles or joint limits.
• Robots whose kinematics differ sharply from the training data may get infeasible predictions.
• Zero-shot performance (a robot the model has never seen) still trails fine-tuning on a little data from that robot.

For fine-tuning, the company reports that some robots needed about 5 minutes of robot data, a Dexmate setup 77 minutes and a modified YAM setup 98 minutes, and that fine-tuning improved language-following on 10 of 12 robots.

The numbers

The headline claim is a ranking. The page says an independent evaluation by Robocurve, a public benefit corporation, on its RoboDojo-RC Tier 1 benchmark, dated 18 October 2026, scored Dream-1 above the next best models, which it names as Opus 5.5 and Gemini 4 Argon. It also claims 16 times lower cost per task attempt than Gemini 4 Argon and speeds 1.7 times faster than GPT-6 Astra. The page does not give the underlying scores in text, and we did not open the Robocurve report, so we cannot verify the ranking.

The generalisation test covered 27 unseen simulated robots, with 300 rollouts per robot design, and four real robots that ran earlier Dream-1 checkpoints that had never trained on them. The exact success percentages appear only in charts on the page.

Dream-13

Dream-1-flash2

Dream-1 listed price per hour (company page) (US$ per hour)

	US$ per hour
Dream-1	3
Dream-1-flash	2

Source: Dream Labs, Dream-1 research page

What it means

For robotics developers and hobbyists. A pay-per-hour API is a different model from downloading weights. It lowers the barrier to experiments on a two-armed robot with parallel grippers, but it also means every action depends on a network call and on Dream Labs' servers staying up and affordable. The page does not state a licence or release open weights, so there is nothing to run offline.

For the field. The seed round is large for a first round. TechFundingNews places it among a series of big robotics-AI raises, including Skild AI at about $1.4 billion in January 2026 and Figure's Series C above $1 billion in September 2025; those figures are that outlet's. Investors are clearly betting that a general-purpose robot "brain" will be a platform business.

Who might lose. Companies whose product is a per-task robot integration could find their niche shrinking if a general model works, though that depends on reliability that has not been shown outside the company's own tests.

What we don't know yet

• Whether the Robocurve ranking holds up. The company cites it; we have not seen the report or the raw scores.
• The model's size and the licence terms. The page gives neither.
• The correct price of the lighter model (see the disagreement above).
• Real-world reliability. The page calls the model a research beta that will make mistakes and has not been validated for safety-critical use; the tasks shown so far are tabletop picking, placing and folding.
• How well the approach carries over to mobile robots and dexterous hands, which are planned but not yet supported.

What to watch next

• Independent tests of Dream-1 by people who do not work for Dream Labs, especially on robots outside its 20-plus hardware partners.
• Publication of the Robocurve evaluation details.
• Support for wheeled robots and dexterous hands, the two extensions the company names.
• Competing hosted robot models from Skild AI, Figure and others, and whether they match the by-the-hour pricing.
• For background on AI agents acting with limited oversight, see our report on the METR and Redwood red-team of OpenAI's agent monitoring.

Sources

• Dream Labs: Dream-1 (primary source)
• SiliconANGLE: Dream Labs raises $90M to help build more intelligent robots with simple video feeds
• TechFundingNews: Ex-Nvidia researchers' Dream Labs raises $90M to give low-cost robots a shared AI brain