Dream Labs announced on October 19, 2026 that it has raised $90 million in seed funding to build foundation models that can control robots made by different manufacturers, according to the company’s announcement and reporting from Pulse 2.0. The round was led by B Capital, with Kindred Ventures, A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC and investor Adrian Aoun also participating. It is one of the largest seed rounds reported this year for a company whose entire product is a brain that other people’s robots can borrow.
The company’s pitch is unusually narrow and unusually ambitious at the same time. Dream Labs is not building a robot. It is building Dream-1, a model trained on video of people doing everyday tasks, and wrapping it in a cloud service called Dream API that turns a robot’s camera feed plus a plain-English instruction into motor commands. Joel Jang, Dream Labs’ co-founder and CEO, framed the bet this way in the company’s announcement: “The hardware is ready. What’s been missing is a brain any robot can use.”
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Dream Labs Banks $90 Million to Build a Robot Brain
The headline number is $90 million, and the headline claim is that Dream-1 is a foundation model usable by “any robot.” That second part deserves a caveat up front: it is the company’s own framing, and independent reporting describes a narrower, though still notable, confirmed scope. What is solidly reported is that Dream-1 is designed to transfer skills across robots from different manufacturers, including machines with grippers, without retraining the model from scratch for each new body.
That distinction matters because robotics has spent the better part of a decade stuck on exactly this problem. A model trained to control one robotic arm historically could not just be copied onto a different arm with a different number of joints, a different camera angle, or a different gripper geometry. Dream Labs is arguing that Dream-1 breaks that pattern, at least for a defined set of tasks and hardware configurations, and it is charging for access to that claim through the Dream API rather than selling hardware of its own.
The $90 million seed round, reported by both Pulse 2.0 and Unite.AI, sits well above the typical seed check for a robotics software startup, and it signals that at least seven named investors believe the cross-embodiment problem is close enough to solved to fund at scale before any product has shipped widely.
Inside Dream-1: How One Model Learns to Run Many Robots
Dream-1 belongs to a category of AI systems researchers call vision-language-action models, or VLA models, which connect what a camera sees, what a person asks for in plain language, and what a robot’s motors actually do. Instead of hand-coding behaviors for each task, a VLA model is trained to predict the next useful action given a video frame and an instruction, then convert that prediction into a stream of motor commands a robot can execute in real time.
Training on Human Video, Not Robot Demonstrations
The detail that separates Dream-1 from older robot-learning approaches is its training data: video of people performing everyday tasks, rather than robot-generated demonstration data collected arm by arm. Collecting robot demonstrations is slow and expensive because someone has to physically operate the machine for every example. Human activity video, by contrast, is abundant. If a model can learn generalizable motion patterns from watching people rather than watching robots, the data bottleneck that has throttled robot learning for years gets a lot smaller.
What “Zero-Shot” Actually Covers
Dream Labs says Dream-1 can control robots it has not previously encountered and can operate machines with grippers without additional training. That is a real, if bounded, claim. It is not the same as saying the model can operate every robot on the market on day one, and nothing in the independently verified reporting supports the broader “any robot” framing used in some coverage of the announcement. Readers should treat the gripper-equipped, previously unseen hardware claim as the confirmed floor, and treat anything beyond that as a company projection rather than an established fact.
A simplified way to picture the interface Dream API exposes looks like this, in conceptual terms rather than an official code sample:
input: {
camera_frame: image,
instruction: "pick up the cup and place it on the shelf"
}
output: {
motor_commands: [joint_1, joint_2, ..., gripper_state]
}That input-to-output shape, camera plus instruction in, motor commands out, is the basic contract most VLA systems share, and it is what lets a robot maker swap in Dream-1 without re-architecting its own control stack.
Dream API Pricing: $3 an Hour for the Flagship Model
Dream Labs has set public pricing for Dream API at $3 per hour for the primary Dream-1 model and $2 per hour for a smaller “flash” variant, according to reporting from Runtime Wire. Charging by the hour, rather than per robot sold or per license, mirrors how cloud AI inference is typically billed, and it reinforces that Dream Labs sees itself as an infrastructure provider rather than a robot manufacturer.
That pricing structure also tells a story about where the company expects its costs to sit. Running a large model that has to generate motor commands in near real time is computationally heavier than typical chatbot inference, which is part of why the flagship tier costs roughly 1.5 times more per hour than the flash tier. For a robot operator running a single arm for an eight-hour shift, the flagship model would cost about $24 a day; running a fleet of machines around the clock would scale that figure quickly, which is likely why Dream Labs is positioning flash as the default option for simpler, lower-stakes tasks.
| Dream Labs Fact | Detail |
|---|---|
| Funding raised | $90 million seed round |
| Date announced | October 19, 2026 |
| Lead investor | B Capital |
| Participating investors | Kindred Ventures, A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC, Adrian Aoun |
| Flagship model | Dream-1 |
| Product | Dream API |
| Primary model price | $3 per hour |
| Flash model price | $2 per hour |
| CEO and co-founder | Joel Jang |
Who Is Behind Dream Labs
Joel Jang is named across the reporting as Dream Labs’ co-founder and chief executive, and his public comment on the raise is the clearest statement of the company’s thesis: the physical hardware for general-purpose robotics has matured faster than the software that tells it what to do. That framing puts Dream Labs in direct contrast with vertically integrated robot makers, which build the arm, the sensors and the control software together. Dream Labs is explicitly betting that splitting those layers, and selling only the software layer through an API, is the faster and cheaper path to scale.
Wing VC, one of the participating investors, published its own rationale for backing the company, describing the wager as one on general robot intelligence rather than on any single robot design, a framing consistent with how Wing VC’s own investment note characterizes the deal.
Why Investors Bet $90 Million on a Seed-Stage Robotics Startup
Seed rounds this large are still rare, but 2026 has produced more of them than any prior year as venture firms chase physical AI rather than software-only AI. Tesla’s Optimus program offers one data point on how fast the hardware side of this market is moving: the company has reported humanoid robot production reaching roughly 1,000 units a week, a pace that would have sounded implausible even two years ago. When robot hardware output scales that quickly, the bottleneck shifts to software that can actually run the machines usefully, which is the gap Dream Labs is explicitly targeting.
Capital is not only flowing toward humanoid platforms, either. Investors have also funded narrower automation bets this year, including a $13 million Series A for crop-pollination robotics, a reminder that venture interest in physical automation extends well beyond warehouse arms and humanoids and into agriculture and other industrial niches. Dream Labs’ pitch, a single reusable control layer that could in principle serve several of these categories at once, is part of why its round drew participation from both generalist and robotics-focused funds.
The Cross-Embodiment Problem Dream Labs Says It Solved
“Cross-embodiment” is the term researchers use for a model’s ability to transfer what it has learned on one robot body to a different robot body. It has been one of the hardest open problems in robot learning because small differences in joint count, arm length, camera placement and gripper design can break a model trained too narrowly. A model that nails pick-and-place on one arm can fail completely on a second arm that looks nearly identical on paper.
Dream Labs’ claim is that training on human activity video, rather than on robot-specific demonstration data, helps Dream-1 learn motion concepts that are less tied to any one machine’s exact geometry. That is a plausible mechanism, and it echoes an approach used by other research groups and a trend captured in Unite.AI’s reporting on the company’s zero-shot scope, but it is also a claim that is easiest to prove in a lab and hardest to prove at scale, across hundreds of real factory floors and warehouses with inconsistent lighting, clutter and hardware wear.
Competitive Landscape: Dream Labs vs. Rival Robot Foundation Model Builders
Dream Labs enters a field that already has well-funded players pursuing overlapping but distinct strategies. Some, like Dream Labs, sell a model and leave hardware to partners. Others build the robot and the model together. A third group treats robotics as one application of a much larger platform. The table below lays out how the main approaches differ.
| Company | Strategy | Hardware Approach | Key Differentiator |
|---|---|---|---|
| Dream Labs | Model and API only | Partners with third-party robot makers | Cloud-billed, cross-embodiment Dream-1 model at $3/hour |
| Physical Intelligence | General-purpose robot policies | Works across multiple embodiments | Focus on broad task generalization, not single hardware line |
| Skild AI | Robot “brain” model | Designed for multiple robot types, including humanoids | General-purpose control layer across embodied systems |
| Figure AI | Vertically integrated | Builds its own humanoid robot | Robot and model developed together in-house |
| 1X Technologies | Vertically integrated | Builds its own humanoid robot | Tight hardware-software coupling for consumer and commercial use |
| Covariant | Manipulation-focused model | Industrial and warehouse robot arms | Deep specialization in picking and sorting tasks |
| Google DeepMind (Gemini Robotics) | Platform ecosystem | Multiple partner robots | Backed by Gemini’s broader vision-language research stack |
| NVIDIA (Isaac GR00T) | Platform ecosystem | Simulation plus partner hardware | Tied to NVIDIA’s simulation and developer tooling |
The split roughly sorts into three camps: model-and-API-first companies like Dream Labs, Physical Intelligence, Skild AI and Covariant; vertically integrated robot builders like Figure AI and 1X Technologies; and platform incumbents like Google DeepMind and NVIDIA, which fold robotics into much larger research and developer ecosystems. Dream Labs is also arriving into a moment when large foundation models more broadly are being released at a rapid clip, including open-weight releases such as Reflection AI’s 501-billion-parameter Beam model, which underscores how much capital and engineering effort is now flowing into foundation models generally, not just the robotics-specific slice of that market.
From Industrial Arms to Foundation Models: A Short History of Robot AI
Industrial robotics has existed commercially since the 1960s, but for most of that history, robots were programmed with fixed, task-specific instructions. A welding arm welded; it did not generalize to packing boxes. The shift toward learning-based robotics accelerated over the past decade as computer vision and language models matured, culminating in research systems like Google’s RT-2, which demonstrated that a model could translate web-scale visual and language knowledge into robot actions rather than requiring task-by-task programming.
Dream Labs’ pitch sits squarely in that lineage, but it pushes the idea further by trying to decouple the model entirely from any single robot manufacturer’s hardware. Where RT-2-style research systems were typically demonstrated on a specific robot platform, Dream-1 is being sold, from day one, as a service meant to run across machines made by different companies. That is a commercial bet as much as a technical one: it only works if enough robot makers are willing to let an outside company’s model sit at the center of their product.
What’s Confirmed and What Isn’t
Several specific figures and claims in Dream Labs’ own announcement have not been independently corroborated by the outside reporting reviewed for this article. The company’s claim that it is working with more than 20 hardware partners appears in its announcement but is not verified by a separate, named source. Likewise, the broad “foundation models for any robot” framing used in some headlines describes a narrower, confirmed scope in the underlying reporting: robots with grippers, including some the model had not previously encountered, rather than literally every robot category.
None of this means the underlying claims are false. It means they have not yet been tested at the scale that would let an outside observer confirm them, which is a normal state of affairs for a company less than a year removed from seed funding. Readers evaluating Dream Labs’ prospects should weight the confirmed technical claims, cross-embodiment transfer for gripper-equipped robots, more heavily than the marketing framing until production deployments generate independent data.
The Technical Hurdles Ahead: Latency, Safety and Sim-to-Real
Running Inference Fast Enough to Matter
Because Dream-1 runs in the cloud rather than on the robot itself, every motor command has to travel from the robot’s camera, to Dream Labs’ data centers, through the model, and back to the robot’s actuators, all fast enough that the robot does not visibly lag or stutter. That round trip is harder to guarantee than it is for a chatbot response, where a half-second of latency is barely noticeable. A half-second of latency on a robot arm moving near a person or a fragile object is a safety problem, not a convenience issue.
The broader AI industry’s compute buildout is relevant context here. Processor makers have been racing to meet demand for exactly this kind of low-latency inference workload; Nvidia’s push into agent-oriented chips, including deals positioning its Vera CPU platform for faster AI agent processing, reflects how much infrastructure investment is being aimed at shaving milliseconds off exactly the kind of real-time inference loop that a cloud-hosted robot model depends on.
Who’s Liable When a Cloud Model Controls a Physical Arm?
Safety and liability questions multiply once a third party’s model, rather than the robot manufacturer’s own software, is making moment-to-moment decisions about where an arm moves. Policymakers have already begun scrutinizing AI safety thresholds in adjacent contexts; proposals such as the discussion around an AI safety funding bar, covered in reporting on an AI safety bill facing industry backlash, suggest regulators are increasingly willing to set explicit thresholds for how much scrutiny advanced AI systems require. Robotics companies whose models directly control physical machinery are likely candidates for similar scrutiny as the sector scales.
Market Impact: What This Means for the Robotics Funding Boom
Dream Labs’ round lands in the middle of a broader surge in capital committed to physical AI, and it arrives alongside tightening supply in the hardware that underpins all of it. Memory makers have reported sharp increases in demand tied to AI workloads broadly; Micron, for instance, has reported HBM4 sales climbing sharply ahead of supplier briefings, with coverage noting sales up 331% heading into an October briefing, a reminder that the compute and memory bottlenecks squeezing large language model training are the same bottlenecks a cloud-hosted robot model like Dream-1 will eventually run into as usage scales.
For robot manufacturers, Dream Labs’ pitch offers a tempting shortcut: skip years of in-house model development and license a control layer instead. For Dream Labs’ well-capitalized rivals, the $90 million round is a signal that the model-and-API strategy has investor backing large enough to compete on compute spend, which is likely to accelerate pricing competition across the sector’s $3-an-hour-and-under tier over the next year.
Predictions: Where Dream Labs and Robot Foundation Models Go From Here
- Expect Dream Labs to publish independent benchmark results or partner case studies within six to twelve months to substantiate its cross-embodiment claims beyond the initial announcement.
- Pricing pressure on hourly robot-API billing is likely, as rival model-first companies such as Physical Intelligence and Skild AI respond with comparable or lower per-hour rates.
- Watch for at least one of Dream Labs’ named hardware partners to go public with a specific deployment, which would meaningfully upgrade the “more than 20 partners” claim from unconfirmed to verified.
- Regulatory attention on AI systems that directly control physical machinery will likely increase, drawing on frameworks already being debated for large language models.
- Vertically integrated competitors like Figure AI and 1X Technologies will likely emphasize safety and reliability in their marketing specifically to differentiate from third-party, cloud-hosted control models.
Frequently Asked Questions
What is Dream Labs?
Dream Labs is a robotics AI company that announced a $90 million seed funding round on October 19, 2026 to build foundation models capable of controlling robots made by different manufacturers.
Who led Dream Labs’ funding round?
B Capital led the round. Kindred Ventures, A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC and investor Adrian Aoun also participated.
What is Dream-1?
Dream-1 is Dream Labs’ flagship model, trained on video of people performing everyday tasks, designed to transfer learned skills across robots from different manufacturers without retraining from scratch for each one.
How much does Dream API cost?
Dream Labs has priced the primary Dream-1 model at $3 per hour and a smaller “flash” model at $0.75 per hour, according to reporting from Runtime Wire.
Can Dream-1 really control a robot it has never seen before?
Dream Labs says the model can control previously unencountered robots, including those with grippers, without additional training. That scope is reported, though broader “any robot” framing in some coverage goes beyond what is independently confirmed.
Who is Dream Labs’ CEO?
Joel Jang is Dream Labs’ co-founder and chief executive. He said of the funding round: “The hardware is ready. What’s been missing is a brain any robot can use.”
Who are Dream Labs’ main competitors?
Named competitors in the robot foundation model space include Physical Intelligence, Skild AI, Figure AI, 1X Technologies, Covariant, Google DeepMind’s Gemini Robotics and NVIDIA’s Isaac GR00T ecosystem.
Is the claim that Dream Labs has more than 20 hardware partners verified?
No. That figure appears in Dream Labs’ own announcement but has not been independently corroborated by other reporting reviewed for this article.
