Dream Labs raises $90M to build one AI brain that controls any robot arm
Dream Labs, a startup founded by former Nvidia robotics researchers, has raised $90 million in seed funding led by B Capital and Kindred Ventures to build Dream-1, an AI model designed to control robot hardware it has never specifically been trained on. Alongside the round, the company launched Dream API, which it describes as the first publicly available API for a robot foundation model. The pitch is blunt: robot hardware has been good enough for years, and the thing actually missing was a brain general enough to run any of it.
How Dream-1 generalizes across hardware
Most robot control systems are trained for one machine performing one task, which is why deploying a new robot arm has traditionally meant months of task-specific programming and data collection. Dream-1 takes a different approach: it's trained primarily on videos of people performing everyday tasks, then learns to predict what should happen next and translate that prediction into motor commands for whatever robot it's controlling. Dream Labs says the model can operate any robot equipped with grippers in a zero-shot setting — meaning no additional training data is needed for a new machine — and currently handles tabletop tasks like picking, placing, and folding laundry.
Co-founder and CEO Joel Jang framed the comparison around rental cars: a driver doesn't relearn how to drive each time they get into an unfamiliar vehicle, they find the controls and apply existing experience. Dream-1 is built to treat a new robot the same way, aiming to get a machine doing useful work within minutes of connection rather than months of integration work. The company is currently working with more than 20 hardware partners whose robots typically cost between $2,000 and $10,000, adapting Dream-1 beyond its baseline two-armed, parallel-gripper configuration for each.
What the API actually gives developers
Dream API connects to a supported robot's camera feed, accepts a plain-English instruction, and returns the motor commands needed to execute it — no robotics-specific machine learning expertise required on the developer's end. For teams that need better performance on a narrow task or unusual hardware, Dream Labs also offers a fine-tuning service: submit a small set of demonstration recordings and receive a customized model version within hours rather than the weeks a from-scratch training run would take. A third product, Robot Playground, lets anyone send commands over a browser to a physical robot in Dream Labs' San Francisco office and watch it attempt the task live — a deliberate choice to let outsiders see real, unedited failures rather than curated demo reels.
Ryan Chan, co-founder of MakerMods, which sells $2,000 bimanual robot arms to Dream Labs' hardware partner network, said the arm was executing attempted tasks it was never explicitly programmed for within minutes of connecting to Dream-1. That's the core value proposition for hardware makers: instead of building and maintaining their own control software, they can ship a physical robot and let Dream Labs' cloud model supply the intelligence layer, improving over time without a firmware update on the customer's end.
Why this is a bet on a specific theory of robotics
Dream Labs is staking its funding on a thesis that's been gaining traction across the robotics industry throughout 2026: that general-purpose “foundation models” for physical action, similar in spirit to how large language models generalized across text tasks, can transfer skills across robot embodiments faster than hand-coded motion planning ever could. The company's co-founders previously worked on Nvidia's GR00T humanoid-robot research program and related World Action Model projects, giving them a direct view into how slowly embodiment-specific robot programming scales compared to model-based approaches trained on broad human demonstration data.
The risk in that bet is generalization quality under real-world conditions. Zero-shot control of unfamiliar hardware is impressive in controlled demonstrations, but industrial and commercial deployments typically demand much higher reliability than a tabletop pick-and-place task requires. Dream Labs' own Robot Playground, which shows live attempts “mistakes included,” is as much an admission of this limitation as it is a transparency feature — the model is good enough to be interesting, not yet good enough to be invisible.
What to watch next
For robot hardware startups, Dream API is worth evaluating now if integration cost is the main obstacle to shipping a useful product — the fine-tuning turnaround of hours rather than weeks is a genuine operational advantage over building control software in-house. For buyers and investors, the signal to track is whether Dream Labs expands task coverage beyond tabletop manipulation into wheeled mobile robots and dexterous hands, which the company has flagged as planned but not yet shipped; that expansion is the real test of whether the zero-shot generalization claim holds up outside a narrow task category.