Dream API Launches as Open-Source Robot Brain API

Reported by Machine Heart

Today, the DreamZero team's Dream Labs officially ended its stealth mode, announcing the completion of a 90,000,000 USD seed round funding, while also launching Dream API: The world's first open API for the public to access a robot brain.

This means that anyone can integrate the Dream API into their own robot, simply by providing natural language instructions, and the Dream API can generate the corresponding motor control commands, allowing the robot to perform tasks. More importantly, it can also adapt to robots that were never seen during training.

Currently, more than 20 hardware manufacturers have joined the collaboration.

This team from NVIDIA proposed the World Action Model (WAM) in February of this year, and WAM is now one of the hottest directions in embodied intelligence.

Behind the Dream API is Dream Labs' first model, Dream-1, which can control robots in real-time and in a closed-loop manner. Developers do not need a background in machine learning to integrate this capability into their own robots, with a cost of only 3 USD per hour. Anyone can try it out at dreamlabs.ai.

Dream-1 promotional video

Dream Labs official website: https://dreamlabs.ai/

This round of funding was led by B Capital, with Kindred Ventures, A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC, and Adrian Aoun participating.

Robots lack a universal brain

A record of 5,000,000 industrial robots are running in factories around the world, but almost every single one is programmed for a single task, and what they learn cannot be transferred to another robot.

At the same time, robot hardware is rapidly becoming cheaper: a dual-arm robotic arm can be as low as about $2000, and increasingly more half-humanoid robots are available for under $15,000. However, these robots almost never come with a ``brain`` at the time of manufacture. Teaching a robot to do one task often takes experts several months; switching to a different task or hardware often requires starting from scratch.

Chatbots have been able to rapidly popularize due to a general model, allowing anyone to use it by simply opening a computer or phone. At Dream Labs, they believe robots also need such a moment. Affordable hardware combined with a model capable of zero-shot generalization, can get the data flywheel spinning: robots becoming useful earlier will lead to more people using them, generating more data, and the model becoming better.

Dream API: The brain of a robot open to everyone

In the past, wanting to have a robot use a foundational model often required training it oneself or deep collaboration with a vendor. Dream API is the first globally open robot foundational model API available to the public: developers can register and connect their robots to the API to use it. The model is deployed across multiple data centers worldwide, allowing robots to respond quickly wherever they are.

Around Dream API, there are two additional services:

• Fine-tuning (FaaS): When encountering unsupported hardware or new tasks, developers and robot manufacturers can submit a small amount of demonstration data, and receive a customized model within a few hours. For a new robot, a single 5-minute trajectory is sufficient to complete the adaptation: on 12 partner robots participating in the Beta test,10 of them saw an improvement in Language following scores.

• Robot Playground: No need to buy a robot, open a browser and give commands to real robots in Dream Labs' San Francisco office, watch their reactions in real time, with no editing involved.

In terms of pricing, besides the Dream-1 at $3 per hour, there is also the faster and lighter Dream-1-flash at $2 per hour.

Guo Yunhao, co-founder and CTO of Dream Labs, said: ``Just like when ChatGPT was first released to the public, robots are finally embracing their first open-to-the-public general brain API. It may not be perfect, but someone always needs to take the first step.''

Dream-1: Learn ``Cross-Entity`` from the Video

Dream-1 can operate on robots from different manufacturers. It trains on videos of people completing everyday tasks, learning to predict what will happen next and then deciding what actions the robot should take. Because it learns from videos rather than from a single machine, the abilities it acquires can be transferred across different hardware: for a dual-arm robot with a gripper, Dream-1 can get up to speed with zero-sample learning right out of the box.

Dream Labs co-founder and CEO Joel Jang gave an analogy: ``You get into a rented car, you don't need to relearn how to drive, find the controls, your past driving experience takes care of the rest. Dream-1 treats new robots the same way. A new robot can do useful work in minutes, not months.''

Why can learning from videos enable cross-ontology learning? Ge Yunhao told Machine Intelligence: ``The world action model simultaneously learns to predict videos and actions, so ``controlling a robot'' and ``imagining what this robot will do next'' become the same problem. From the perspective of the top camera, any robot is simply a moving shape in the picture, and the video is a shared representation for all robots. As long as the model can predict what a new robot looks like when completing a task, it can generate the actions that will lead the robot to that state.

Robots and tasks that have never been seen before can also get up to speed with zero samples

The team evaluated 27 types of Dream-1 in simulation From robots that have never been seen before, including Unitree R1, Baxter, dual-arm UR5e, ALOHA, etc., each ran 300 rollouts. Taking YAM's performance as 100%, 24 of them reached over half, with the highest dual-arm PiPER reaching 91.2%.

On real hardware, Dream-1 also achieved zero-sample control of MakerMods' Metal Arm and Maker Arm, Seeed Studio's reBot B601-DM, Anvil Robotics' OpenYAM, and Dexmate's Vega etc. partner robots, completing tasks such as folding towels, closing drawers, hanging earphones on hooks, and ``putting fruits that are not kiwifruit into the basket''.

Dream-1 Zero-shot Control New Robot

It can also complete unseen tasks in unfamiliar environments. Beta test partners run Dream-1 autonomously at 9 global sites, including Shanghai's AGIBOT、Shenzhen's BitRobot, KAIST in Seoul, Hugging Face in Paris, Scale AI in San Francisco, etc., tasks range from putting a screwdriver into a box, folding a box, to sorting tools into storage boxes.

Dream API

Currently, Dream-1 mainly handles desktop tasks, such as grasping and placing objects, folding clothes. The dual-arm parallel gripper robot can operate without additional data; for over 20 hardware partners, Dream Labs will further optimize the model for each robot, with these robots' prices ranging from 2000 to 10000 USD.

Can large models directly control robots, and do robots need a brain?

Recently, general large models like GPT-6 Astra have been directly connected to robotic arms to control robots to work, and many people have started to worry: will specialized robot base models be replaced by large models?

In the RoboDojo-RC Tier 1 evaluation released by the independent third-party evaluation institution Robocurve, Dream-1 achieved an average score of 59% per task attempt, the first model to exceed 50%, higher than Claude Opus 5.5 (48%), Gemini 4 Argon (48%), GPT-6 Astra (47%), GPT-6.1 Sol (42%), and other latest large models, as well as MolmoAct2 (19%) and GR00T N1.7 (10%) and other robot models.

Not only is the score higher, Dream-1 is also cheaper and faster: the cost per task attempt is about 0.15 USD, compared to the second-highest scoring Gemini 4 Argon, the score is 1.2 times higher, but the cost is about 16 times lower. The average time is 1.4 minutes, compared to the fastest large model GPT-6 Astra (ultrafast), the score is 1.3 times higher, and the speed is 1.7 times faster. Each model configuration has been tested 120 times in the real world.

NVIDIA GR00T Four Core Members

The four co-founders all come from the NVIDIA GR00T robot team:

• Joel Jang (Co-founder and CEO): NVIDIA Responsible for the world model of GR00T, led LAPA, DreamGen, DreamDojo and DreamZero;

• Ge Yunhao (Co-founder and CTO): NVIDIA Responsible for the video model of GR00T, participated in building the NVIDIA Cosmos world base model series (Predict, Transfer, Policy), with a career spanning Stanford, USC, Google and Microsoft;

• Scott Reed (Co-founder and Chief Scientist): Founding member of NVIDIA GEAR Lab, led the GR00T N1 to N1.7 series models; previously worked at DeepMind for 8 years, created Gato and Genie, with over 158000 citations on Google Scholar;

• You Liang Tan（Co-founder and Chief Robot Officer）: NVIDIA GR00T Open Source Lead, participated in building ROS 2, Octo, SERL, EgoScale. Full-stack robot expert.

In February of this year, this team released DreamZero while at NVIDIA, for the first time explicitly proposing the World Action Model (World Action Model, WAM). In the following half year, WAM quickly became one of the hottest routes in embodied intelligence.

At that time, the biggest surprise for the team was: DreamZero was able to zero-shot generalize to tasks it had never seen before with just 30 minutes of YAM robot data for post-training, whereas in the past this often required hundreds or even thousands of hours of specific domain data. If 30 minutes is enough, the robot's 'no data, no use; no use, no data' dead loop could potentially be broken. The team then decided to start a company. Six months later, Dream-1 was released.

The lead investor, B Capital Partner (General Partner) Daisy Cai, stated: “A functional robotic arm has already dropped in price from that of a luxury car to that of a high-end laptop. What it now lacks is intelligence. This team has done some pioneering work in the field of robot learning at NVIDIA, and now anyone who gets Dream API can build upon it.”

Limitations and Next Steps

Dream-1 is currently a research Beta version, and the team has also stated its limitations:

• Performance depends on the quality of the inverse kinematics solver for each robot;

• It only determines where the gripper ends up, without controlling how the arm reaches there, which may cause issues near obstacles or at joint limits;

• When facing or having encountered a body that is very different from the robots it has seen, it may predict unfeasible actions.

Joel Jang said: ``We don't show everyone a edited demo, but instead give everyone a model that can be tested themselves, truly usable, along with its errors.''

There is still a gap between zero-shot and one-time adaptation. Ge Yunhao said: ``We really want to eliminate even these 5 minutes. The gap between zero-shot and one-time adaptation is what we are focusing on most in the next step.``

Next step, Dream-1 will support wheeled mobile robots and dexterous hands. This round of financing will mainly be used for computing power to expand the tasks Dream-1 can handle and the types of robots, and to support more users.

In a world of robots where hardware is getting cheaper yet still lacks the 'brain,' how far can a model that can directly take over an unfamiliar robot go? Dream-1 has submitted its first testable answer.

Finally, here's a hidden gem:

Reference link:

https://dreamlabs.ai/

https://dreamlabs.ai/research/dream-1

© THE END 

For republication, please contact this official account for authorization

For submissions or media inquiries: [email protected]