Dream Labs raises $90m to put everyday robotics within reach
A robot arm can now cost a few thousand dollars. Getting it to fold a shirt or put away tools remains a different proposition. Dream Labs has raised $90 million to narrow that gap, with an ambition it describes as “robotics within reach”: making machines easier to use and accessible to people beyond specialist laboratories.
Its Robot Playground offers a playful glimpse of what that could mean. Visitors can dream up instructions and watch a robot attempt them. One visitor’s instruction was to turn bowls upside down to resemble a drumkit, then play them with spoons. The request captures the fun of testing how far a machine can follow an imaginative idea, even when success is far from assured.
The same flexibility could have practical value in industry. A machine that can interpret a new instruction might be easier to adapt to sorting components, packing products or rearranging a workbench. The ambition is to let people request a new task without programming every movement, although an entertaining experiment is still some distance from dependable workplace automation.
The San Francisco-based company is still in the early stages of turning that ambition into a dependable product, but it has already secured substantial backing. Its $90 million seed round, announced on 19 October 2026, was led by B Capital alongside Kindred Ventures, with participation from investors including A.Capital, Unusual Ventures, SV Angel and Wing VC. Dream Labs says the money will go primarily towards computing resources as it expands capabilities and serves more users. Its cloud-based approach means computing is an ongoing operating requirement as well as a research expense: customers’ robots rely on models running in data centres.
Alongside the financing, it launched Dream API, an application programming interface through which developers connect robots to its Dream-1 model. Visual inputs and instructions are used to generate actions. The proposition is straightforward: manufacturers build the bodies; Dream Labs supplies a service that helps control them.
Who founded Dream Labs?
Founded in 2026, Dream Labs brings together four co-founders from NVIDIA’s robotics programme. Chief executive Joel Jang led GR00T’s world-model research. Chief technology officer Yunhao Ge worked on its video models and NVIDIA’s Cosmos technology. Chief scientist Scott Reed led DeepMind’s Gato, which used one model for language, games and robot control, before leading GR00T N1 at NVIDIA. You Liang Tan brings experience in ROS 2, the Robot Operating System framework, and NVIDIA’s open-source robotics efforts.
From DreamZero research to Dream-1
The team’s research foundation includes DreamZero, released at NVIDIA in February 2026. It introduced a World Action Model: a system that jointly predicts future video and robot actions.
The idea is that recognising a cup is only part of picking it up. A robot also needs to anticipate how its movements will change the scene. A model trained to predict those changes can learn something about physical interactions, alongside the movements needed to produce them.
In experiments reported in the DreamZero paper, the researchers found more than twice the performance on new tasks and environments compared with the vision-language-action models they tested against. These models connect visual and language understanding with physical actions. DreamZero also adapted to a new robot using about 30 minutes of interaction data.
The results were specific to those experiments but suggested a way to reduce the expensive process of collecting extensive training data for every new machine.
Dream-1 takes that research direction into a commercial robot foundation model: an AI system intended to provide reusable capabilities across tasks and machines. Dream Labs says it learns from video of people performing everyday activities, helping knowledge travel beyond an individual robot’s training experience.
How Dream-1 works across different robots
For unfamiliar hardware, Dream-1 can specify where grippers should move and whether they should open or close. The robot’s own motion software translates those targets into movements of its joints. That provides a shared interface across differently built machines, although performance still depends on the hardware and its software.
The current baseline is two-armed robots with parallel grippers, whose fingers move towards or away from each other. Broader mobile and hand capabilities remain development goals, although Dream Labs has demonstrated adapted hand configurations.
The strategy recalls Android’s role in smartphones: different manufacturers building around shared software.
Dream API: capabilities, pricing and limitations
Dream Labs says it works with more than 20 hardware partners whose supported machines cost roughly $2,000–$10,000. Demonstrated tasks include picking and placing objects and folding clothing.
Dream Labs lists Dream-1 at $3 per hour of model use, with a lighter Dream-1-flash version at $2. For unsupported hardware or additional capabilities, a fine-tuning service accepts demonstration data to produce a tailored model. However, Dream-1 remains a research beta, unvalidated for safety-critical use. Its documentation acknowledges limitations around obstacles, joint limits and unfamiliar robot designs, notably.
Dream Labs’ competitors in robot AI
Dream Labs is entering an active field. Physical Intelligence’s π0.7, introduced in April 2026, is designed to combine learned skills and work across robot platforms, scenes and tasks. Its training includes robot experience and human video, underlining that learning from people is also being pursued by competitors.
Google DeepMind’s Gemini Robotics 2 includes models for reasoning and physical control, with an on-device version that runs locally. That offers a different deployment option for applications sensitive to connectivity and network delays. NVIDIA’s GR00T models provide another foundation developers can build upon.
Dream Labs seeks to differentiate itself by reducing the expertise needed to put robots to work. The $90 million gives it resources to pursue that goal. The journey from a spoon-wielding drumkit experiment to reliable everyday assistance will depend on how consistently its robots can turn instructions into completed tasks.
Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.