On October 19, robotics AI startup Dream Labs announced the close of a USD 90 million seed round alongside the launch of its robot foundation model, Dream-1, and a public API.

Raising USD 90 Million in Seed Funding, Dream Labs Unveils New Model

On October 19, robotics AI company Dream Labs announced it has closed a USD 90 million seed funding round, simultaneously launching its robotics foundation model, Dream-1, and a public API. The round was led by B Capital, with Kindred Ventures as co-lead. Other participating firms include A.Capital, Unusual Ventures, SV Angel, Wing VC, and Scribble VC, alongside individual investors such as Hugging Face CTO Thomas Wolf and Adrian Aoun.

This marks Dream Labs' public debut. The company plans to allocate the funding toward compute infrastructure, model capability expansion, and developer ecosystem development, advancing robotics foundation models from research into broader real-world deployment.

USD 90 Million Seed Round: Funding Directed Toward Compute and Developer Ecosystem

The investor roster spans early-stage technology venture firms and industry-focused investors. B Capital led the round, with Kindred Ventures as co-lead. Other participants include A.Capital, Unusual Ventures, SV Angel, Wing VC, and Scribble VC, along with individual investors Thomas Wolf and Adrian Aoun.

Dream Labs plans to allocate the funding across three primary focus areas.

• Compute infrastructure: Supporting the training and iteration of robotics foundation models.
• Model expansion: Increasing the variety of tasks the model can handle and the range of compatible robot hardware.
• Developer ecosystem: Broadening the model's adoption and enabling more developers to access robotic intelligence capabilities.

This funding reflects Dream Labs' chosen commercialization path for robotics foundation models. Rather than selling robot hardware directly, the company aims to provide general-purpose intelligence across diverse robots via its models and APIs, eliminating the need for developers to repeatedly train and build for specific hardware platforms.

However, public information has yet to disclose the valuation associated with this funding round or the specific allocation of the raised capital.

Dream-1 opens its API, bringing the model from the lab to robotics developers

Along with the funding announcement, Dream Labs officially launched Dream-1, a robotics model based on a world action model (WAM), and released a public API.

Developers can feed robot camera streams and natural language instructions into the Dream API to generate corresponding robot action commands. Dream-1 currently targets dual-arm robots equipped with parallel grippers, supporting desktop manipulation tasks such as grasping, placing, and folding.

Dream-1 supports desktop manipulation tasks such as grasping, placing, and folding for dual-arm robots equipped with parallel grippers.

Regarding pricing, Dream-1 costs USD 3 per hour to run, while its lightweight variant, Dream-1-flash, costs USD 2 per hour. For specific tasks the base model cannot execute directly, developers can use a Fine-tuning-as-a-Service (FaaS) offering to adapt the model using data from the target robot.

Dream Labs states that the model has been tested on multiple physical robots, showcasing results across 27 robot morphologies not included in its training data. The company also notes it is currently working with more than 20 hardware partners.

These developments reflect the product ecosystem Dream Labs aims to build: centered on a robotics foundation model, it leverages APIs, hardware adaptation, and fine-tuning services to lower the barrier for developers integrating robotic intelligence.

However, cross-robot transfer remains constrained by hardware structure, kinematics, and control systems. Furthermore, strong performance on test tasks does not equate to stable deployment capabilities in complex, real-world scenarios.

From DreamZero to Dream-1: The Team Continues to Back World Action Models

Dream Labs' technical lineage traces back to DreamZero, released in February 2026. This research explores world action models, aiming to integrate physical world understanding with action prediction. This enables the model to not only comprehend task instructions but also predict the potential state changes following a robot's executed actions.

Dream Labs co-founders Joel Jang, Ge Yunhao, and Scott Reed bring research experience from NVIDIA, Google DeepMind, and top academic institutions.

The Dream Labs team comprises members with prior experience at NVIDIA, Google DeepMind, Ai2, and the Toyota Research Institute, alongside academic affiliations with Stanford University and the University of California, Berkeley.

Co-founder and CEO Joel Jang previously led world model research for Project GR00T at the NVIDIA GEAR Lab and co-led the DreamZero project. Through his work on DreamGen, DreamDojo, and DreamZero, he has continuously explored how to translate video-based experiences into training data for robots. His research trajectory highlights the team's technical accumulation in transitioning from video data to action control prior to founding the company.

Co-founder and CTO Yunhao Ge previously contributed to the research on NVIDIA's Cosmos series of world foundation models before transitioning to GR00T-related work. He holds a Ph.D. in Computer Science from the University of Southern California and was a visiting scholar at the Stanford Vision and Learning Lab. His personal website highlights Cosmos, Cosmos Policy, and DreamZero as his signature research, reflecting a technical focus on video-based world models and their real-world robotic deployment.

Co-founder and Chief Scientist Scott Reed has long focused on generalist agents and world models. He is the lead author of the DeepMind paper on Gato, a generalist agent, and contributed to the development of the world model Genie. Previously a Senior Staff Research Scientist at DeepMind, he later became a Principal Research Scientist at NVIDIA, where he helped establish the GEAR Lab. His foundational research in generalist agents, world models, and robotic learning directly underpins this venture.

The DreamZero research explores combining physical world understanding with action prediction to anticipate state changes during robot execution.

Following this funding round, the company plans to expand the range of tasks and robot embodiments supported by its models. It also aims to refine its development toolchain, spanning model training and simulation validation to on-robot deployment.

For robotics foundation model companies, model capability represents only one facet of commercialization. Equally critical to building a sustainable business is reducing hardware integration costs, improving robustness across diverse tasks, and enabling developers to consistently deploy these models in real-world products.

The USD 90 million seed round provides Dream Labs with the capital to scale its compute infrastructure, expand its model capabilities, and grow its developer ecosystem. Moving forward, the critical question is whether Dream-1 can transition from its current testing and partnerships to stable deployment, and whether its API services can develop into a sustainable business model.