Dream Labs Opens First Public Robot Brain API, Backed by $90M Seed Round

Dream Labs exited stealth on Monday with a $90 million seed round and the public launch of Dream API — the first cloud inference service designed to give any robot a working brain. The San Francisco startup, founded by four former NVIDIA GR00T team leaders, opened developer access to Dream-1, a World Action Model capable of controlling unfamiliar robots in real time without robot-specific training data. B Capital and Kindred Ventures led the round, with A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC, and investor Adrian Aoun participating.

The Robot Brain Cold-Start Problem

Five million industrial robots are operating globally, but nearly every one of them runs fixed code written for a single task on a single machine. Entry-level bimanual robotic arms now cost as little as $2,000, and humanoid-adjacent platforms are available for under $15,000 — hardware prices that have dropped well within reach of a warehouse startup or a university lab. The bottleneck is no longer mechanical.

Teaching a robot a new task has historically required months of engineer time and hundreds of hours of robot-specific teleoperation data, with learned policies locked to that specific platform. Move to a different arm and the process restarts from scratch. This cold-start trap has prevented inexpensive hardware from producing useful autonomous robots at scale: a robot only becomes valuable once deployed, but deployment requires expertise most buyers do not have.

A World Action Model, Not a VLA

Dream-1 is a World Action Model — a class of robot policy architecture first formally proposed in DreamZero, a February 2026 paper from NVIDIA's GEAR Lab co-authored by three of Dream Labs' four founders.

Most robot foundation models are Vision-Language-Action models, or VLAs. A VLA predicts robot actions directly from visual observations and language instructions, outputting joint angles or velocity commands in a robot-specific action space. The architecture trains well on data from the target robot, but it binds the model's learned representations to a specific morphology — change the arm geometry and the policy typically fails.

Dream-1 changes the prediction target. It is an autoregressive diffusion transformer trained to jointly produce two outputs simultaneously: future video frames from the robot's head-mounted camera, and end-effector pose sequences describing where the gripper should move. The video prediction component is the key to cross-embodiment transfer. From a head camera's perspective, any bimanual robot is shapes moving across a workspace — and those shapes look functionally the same regardless of whether the underlying arm has four joints or six, or whether it was manufactured by Unitree, UR, or Seeed Studio.

According to Dream Labs' research post, a World Action Model "learns to predict video and actions together, so controlling a robot and imagining what that robot will do are the same problem." Because every robot appears to the head camera as shapes moving through a scene, a single model gains a common representational basis for controlling hardware it has never seen before.

Dream-1 runs at approximately 7 Hz in closed-loop operation. Each API call returns 1.6 seconds of future actions as end-effector pose sequences, which the robot's own inverse kinematics solver translates to motor commands.

Read more: NVIDIA and KAIST cut per-task robot foundation model RL training time by up to 68.5%

Dream API: Infrastructure at $3 an Hour

Dream Labs is selling compute, not robots. Dream-1 is priced at $3 per hour of inference; Dream-1-flash, a lighter and faster variant, costs $2 per hour. At approximately 2,250 API calls possible per hour, each individual action step costs around $0.0013 — comparable in scale to a token from a small language model API. The pricing positions the company explicitly as infrastructure: a robot hardware OEM can outsource the intelligence layer entirely rather than building AI training capability in-house.

The API accepts head-camera images and natural language instructions via a Python client and returns motor commands without requiring machine-learning expertise on the developer's side. A fine-tuning service — FaaS — handles robots needing additional adaptation: five minutes of teleoperation demonstration data is sufficient to onboard most new hardware, according to the company, with 10 of 12 beta hardware partners seeing improved language-following scores. For morphologically complex systems such as the Sharpa dexterous hand, 77 minutes of demonstration data was required.

CEO Joel Jang described the proposition in the official press release: "The hardware is ready. What's been missing is a brain any robot can use. When you get into a rental car, you don't relearn how to drive. You find the controls and your past driving experience does the rest. Dream-1 approaches a new robot the same way, so a new robot can do useful work in minutes, not months." B Capital General Partner Daisy Cai framed the investment thesis directly: "A capable robot arm has gone from the price of a luxury car to the price of a high-end laptop. What it needs now is intelligence."

What the Robocurve Numbers Actually Show

The most concrete external evidence for Dream-1's capability comes from RoboDojo-RC Tier 1, a real-robot manipulation benchmark administered by Robocurve, an independent public benefit corporation. Each model configuration was evaluated in 120 trials on real hardware. Dream-1 scored 59% — the first model on the benchmark to clear the 50% threshold. Among frontier models applied directly to robotic hardware, Gemini 4 Argon scored 48%, Claude Opus 5.5 scored 38%, GPT-6 Astra scored 43%, and GPT-6.1 Sol scored 42%. Among specialist robot models, NVIDIA's GR00T N1.7 scored 10%, and MolmoAct2 scored 19%.

Benchmark results deserve scrutiny. RoboDojo-RC Tier 1 measures language-instruction following in tabletop manipulation tasks; it is not a general measure of robotic capability, and evaluation conditions for cloud API-based models may differ from locally deployed models. Dream Labs' cost comparison placing Dream-1 at approximately 16 times lower per-task cost than Gemini 4 Argon is a company-derived figure corroborated by Robocurve's own cost-per-attempt analysis. Robocurve's methodology has not yet been independently audited — the results arrived on launch day.

Even with those caveats, the pattern carries weight. A purpose-trained robot model consistently outscores frontier models orders of magnitude larger but not trained specifically for physical manipulation. The gap between GR00T N1.7 at 10% and Dream-1 at 59% is difficult to dismiss as benchmark noise given they originate from the same research lineage — it more likely reflects the architectural advantage of predicting video-frame futures over platform-specific joint trajectories.

Zero-Shot Across 27 Robots, Real and Simulated

In simulation, the team tested Dream-1 on 27 robot designs not encountered during training, running 300 rollouts per robot. Twenty-four of the 27 reached at least half the performance of the training reference platform; the highest-performing unseen design, a bimanual PiPER configuration, reached 91.2% of baseline. On real hardware, Dream-1 controlled robots from MakerMods, Seeed Studio, Anvil Robotics, and Dexmate without any robot-specific training data, completing tasks including folding towels, closing drawers, and executing language instructions such as "put the fruit that is not a kiwi into the basket."

Beta partner deployments ran autonomously across twelve global sites: AGIBOT in Shanghai, BitRobot in Shenzhen, KAIST in Seoul, Hugging Face in Paris, and Scale AI in San Francisco, among others. The cross-embodiment approach is most reliable for dual-arm parallel-gripper robots morphologically similar to the training distribution; the company's documentation acknowledges that substantially different configurations may produce less accurate predictions.

Read more: Meta FAIR establishes first scaling laws for multi-embodiment robot world models

Research Beta: Limitations the Company Is Transparent About

Dream Labs ships Dream-1 with an explicit list of known failure modes. The model predicts end-effector destination poses, not joint trajectories — a separate inverse kinematics solver on the robot translates EEF targets to motor commands, and if that solver is imprecise or the arm is near joint limits, instructions may produce unreachable configurations. Because the model operates in EEF pose space, it has no awareness of obstacles in the arm's path, creating collision and infeasibility risks in cluttered workspaces. Robots with substantially different end-effectors — multi-fingered dexterous hands, mobile bases, non-standard tool configurations — remain partially unsupported.

Dream-1 is a research beta explicitly not validated for safety-critical applications. Jang said in the press release that rather than releasing polished demos, the company is "giving people a working model they can test themselves, mistakes included."

Next: Mobile Robots, Dexterous Hands, and the Data Flywheel

Dream Labs has disclosed two near-term expansion targets: wheeled mobile manipulators and dexterous multi-fingered hands. The company's research post notes that closing the gap between zero-shot and one-shot performance — specifically, the distance between what Dream-1 can do with no robot-specific data and what it achieves after five minutes of demonstrations — is among the team's main ongoing goals. The $90 million round is expected to fund primarily compute, scaling Dream-1 to broader task types and robot morphologies.

The structural question Dream-1 poses is whether robot hardware manufacturers will build their own AI intelligence layers or outsource them to cloud APIs — a dynamic that mirrors how LLM APIs restructured the software industry's relationship with language models. If a cloud robot-brain API can generalize across hardware at production scale, OEMs compete on kinematics, cost, and durability rather than proprietary AI training infrastructure. Whether Dream-1 delivers that promise, or whether Day 1 benchmark results and beta deployments represent conditions too controlled to generalize, is what the next year of real-world deployment data will answer.