Dream Labs: Building the Brain for Everyday Robots

Dream Labs co-founders, from left to right: You Liang Tan, Scott Reed, Joel Jang, Yunhao Ge

Our Investment in Dream Labs

Picture a $2,000 robot arm on a workbench. It has two grippers, a camera, and motors precise enough to fold a shirt. It's plugged in and ready, and it does nothing. Nobody has spent months teaching it a task, and anything it did learn would stay locked inside that one machine.

Everyday robots have been stuck in that gap between capability and intelligence for years. Dream Labs was created to close it.

Today, we're super excited to announce that Kindred Ventures invested in Dream Labs' $90M seed round alongside B Capital who led the round with participation from investment firms, A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC, and Thomas Wolf (CTO of Hugging Face), Adrian Aoun, and other angels. As part of the investment, our founder, Steve Jang will join the company's board. Today, alongside the financing, the company is launching Dream API, the first publicly available API for a robot foundation model. You give it a robot's camera feed and a plain-English instruction, and it returns motor commands, even for robots it has never seen before.

At Kindred, we've been looking for founders building true world models, foundation models that go beyond language and learn how the physical world actually works. Dream Labs is that thesis, just made concrete, shipping as a developer-facing product on day one.

The Hardware Is Ready. The Brain Not So Much

There have never been more high-performing robots. The International Federation of Robotics reports that the number of industrial robots in operation worldwide grew 9% in 2025 to a record 5 million, with more than 600,000 new units installed in a single year. Yet nearly every one of them is programmed for a single task. Its knowledge can't move to the machine next to it. A new task or a new robot often means months of specialist work, starting from zero.

Meanwhile, the cost curve has collapsed. A capable robot arm used to cost as much as a luxury car. Now it costs about as much as a high-end laptop. The robots Dream Labs supports today cost roughly $2,000 to $10,000. But they ship with no built-in intelligence.

The bottleneck has moved. It's no longer metal, motors, or margins. It's the brain.

We've watched this pattern play out across every computing platform we've invested through. PCs needed an operating system. Smartphones needed an app platform. Generative AI broke out once foundation models were available behind an API. Robotics has reached the same moment: abundant hardware waiting for a general-purpose intelligence layer any machine can plug into.

From Describing the World to Imagining It

For the past few years, the dominant recipe for generalist robot policies has been the vision-language-action model, or VLA. You take a pretrained vision-language model and fine-tune it to output robot actions. That approach carried the field a long way, including open models like NVIDIA's GR00T N1, which Joel helped build. But it has a ceiling. In NVIDIA's own framing, a vision-language backbone learns to describe the world, not to predict how it evolves. A robot needs exactly that prediction whenever a task depends on how a scene will change.

A world action model (WAM) begins somewhere else. WAMs (a term coined by Dream Labs' co-founder Joel Jang) learn both how the world is likely to change and what actions a robot can take to shape that change. Video is the core training signal, the densest, most abundant record we have of how the physical world moves over time. The model imagines the next state and the motion and dexterity that gets it there, together.

A robot that can imagine the next second can act in it.

World Action Models in Action

The Dream Labs founders created this category. In February 2026, Joel and his colleagues at NVIDIA's GEAR Lab released DreamZero, a 14-billion-parameter world action model built on a pretrained video diffusion backbone. The results reset expectations for what a robot policy could do:

• More than 2x better generalization to new tasks and environments compared with state-of-the-art VLAs in real-robot experiments (arXiv).
• Real-time control at 7Hz from a 14B video diffusion model, achieved through model and systems optimization (arXiv).
• Learning by watching. 10-20 minutes of video-only demonstrations from humans or other robots improved performance on unseen tasks by more than 42% (arXiv).
• New bodies in 30 minutes. DreamZero adapted to an entirely new robot with only 30 minutes of play data while keeping its zero-shot abilities (arXiv).
• Top of the leaderboard. On RoboArena, which tests real robots on arbitrary natural-language commands, DreamZero reached 1750 Elo in April 2026, ahead of many other VLA-focused models.

That work rests on a longer research arc. DreamGen showed that video world models could generate synthetic "neural trajectories" that teach robots new verbs in new environments. DreamDojo pretrained a robot world model on 44,000 hours of first-person human video, the largest dataset of its kind, and its learned simulator predicted real-world policy success with a 0.995 correlation . The field has moved quickly behind these ideas. Jim Fan's "Robotics' End Game" talk has been summarized as "VLAs are dead, long live World Action Models" and public WAM efforts now include work from Ant Group, Rhoda AI, Sereact and Mimic Robotics.

Physics Generalizes

The core insight behind Dream Labs is simple and profound: physics generalizes better than semantics. A cup falls the same way in a kitchen as it does in a lab. Cloth folds the same way regardless of who manufactured the gripper. A model that has learned dynamics carries that knowledge into scenes and motions it never saw in training.

That one shift opens three doors that have kept robotics narrow.

Data scales with people, not robot fleets. Robot data is scarce, expensive, and slow to collect, and VLAs often need many near-identical demonstrations of the same task. World action models can learn from varied, messy interaction data, including video of people doing ordinary things. We made the same bet with our investment in Mecka, where we wrote that egocentric human capture is a road that "scales with people rather than with just robot fleets." Dream-1 is trained on video of people performing everyday tasks, and it turns that abundance into robot skill.

Embodiment stops being a wall. Because Dream-1 learns from video and not from one machine's motor logs, its capabilities transfer across hardware. Joel puts it best: "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." Out of the box, Dream-1 controls any robot with grippers zero-shot, and it needs no additional data to run a two-armed robot with parallel grippers. For each of its 20+ hardware partners, Dream Labs refines the model further so every robot performs at its best.

Imagination becomes a tool. A world action model predicts what will happen next, not just what to do. That makes it both a policy and a learned simulator, a foundation for evaluation, planning, and self-improvement that a pure action model can't offer. DreamDojo already showed that planning inside a world model improved real-world success on a fruit-packing task by 17%.

Put together, these are the ingredients for a scaling curve that looks far more like language models than like traditional robotics. More video and more compute produce better physical priors, and those priors flow to every robot connected to the platform.

Robot Intelligence, One API Call Away

Most robot foundation models have lived behind closed doors or shipped as research weights that need a robotics lab to run. Dream Labs is doing something different. It's putting general-purpose robot intelligence directly in developers' hands:

• Dream API. Sign up, connect a robot with grippers, describe the task in plain English, and the robot carries it out. Inference runs in data centers around the world, so robots respond quickly wherever they are.
• Fine-tuning as a Service. For unsupported hardware or new capabilities, developers and robot makers submit a small batch of demonstrations and receive a tailored model within hours.
• Robot Playground. Anyone with a browser can type a command to a real robot in Dream Labs' San Francisco office and watch it respond live and unedited, no robot purchase required.

That last one matters more than it might seem. Robotics has been flooded with polished, heavily edited demo reels. As Joel says, "instead of edited demos, we're giving people a working model they can test themselves, mistakes included." Faster deployments means more interactions earlier, and every interaction becomes a real-world signal for the team.

Dream Labs also sits in a horizontal enabler position in the ecosystem: the intelligence layer in the physical AI stack, not a competitor to the robot makers. Ryan Chan, co-founder and CEO of MakerMods, which sells $2,000 bimanual arms, describes his robot being "up and running in a few minutes and attempting tasks we never programmed." Hardware makers get to focus on building better hardware, and their customers get intelligence that keeps improving. Every new robot, developer, and deployment widens what the model has seen. That flywheel compounds.

Real-World Robot Rollouts

The Team That Dreamed Up The World Action Model

We're magnetically drawn to founders who invent the technique, then insist on commercializing it into something people can use, and want to build an incredibly fast-moving juggernaut of a company around this vision. Joel and his co-founders led world-model research for NVIDIA's GR00T program, and they are the researchers behind LAPA, DreamGen, DreamDojo, and DreamZero, working closely with our friends Jim Fan and Mingyu Liu. They didn't adopt the world-action-model paradigm. They defined it and even coined the name.

Joel Jang, CEO, (no relation to me :)) came to robotics through language-model research at KAIST and doctoral work at the University of Washington. At NVIDIA's GEAR lab, he led GR00T's world-model effort, with work spanning LAPA, DreamGen, DreamDojo, and DreamZero.

While at Nvidia, Joel coined the term "World Action Models." His work tackles a practical question: how can a robot turn what it learns from watching people into actions of its own? Physical knowledge can transfer from human video, but putting those lessons to work requires adapting them to the robot's body and viewpoint.

Yunhao Ge, CTO, earned his computer science PhD at USC and researched computer vision at Stanford. At NVIDIA, he led GR00T's video-model work and helped build Cosmos across Predict, Transfer, and Policy.

He co-authored Cosmos Policy, which uses robot demonstrations to adapt a pretrained video model into a policy that generates actions and predicts future states.

Scott Reed, Chief Scientist, earned his PhD at the University of Michigan and spent eight years at DeepMind. He developed Gato's concept, wrote its initial prototype, and led the project. Gato used one set of weights for language, images, games, and robot control.

He also advised early Genie research before joining NVIDIA GEAR as a founding member. There, he led GR00T N1, an open model combining vision, language, and action generation for humanoid robots.

You Liang Tian, Chief Product Officer, trained in mechanical engineering at Nanyang Technological University and computer science at Georgia Tech. At Open Robotics, he helped build ROS 2 and Open-RMF, software for coordinating robot fleets.

After robot-learning research at Berkeley's RAIL Lab, he led GR00T's open-source efforts at NVIDIA. His contributions to Octo, SERL, and EgoScale span robot learning and the practical tools other roboticists use to build on it.

When we first met them, what struck us most is how clearly the team understands the distance between a breakthrough paper and a product. A common challenge for teams we see coming from well-funded multi-trillion dollar companies: the great difficulty in bringing research and model development into the commercial universe where powerful learning loops occur. The DreamZero research project showed what a world action model could do in a lab. Now, Dream Labs, the startup, is building everything needed beyond that: inference infrastructure that keeps robots responsive around the world, a fine-tuning pipeline that brings new hardware online in hours, partnerships with more than 20 robot makers, and the confidence to let anyone watch the model work, and fail, in real time. Very few teams can do the research. Fewer still will do the work of making it reliable for diverse and opinionated builders on the internet. That gap and the intrinsic ambition and mission-driven speed are where and how durable companies get built in the AI space.

Why Physical AI Needs This Now

When we announced our latest funds, we wrote that the next era of technology starts with novel intelligence, and that it will happen in physical AI, "where consumer robots both serve as companions and tools, where industrial robots transform and expand factories, datacenters, and manufacturing plants." Our experiences in autonomous vehicles, specialized robotics such as neurosurgical microrobots and character robots, and generative media have led us to an increasing belief in a deep learning approach using transformer and diffusion models using video, real-to-sim-to-real, and robotic hardware data in the training pipeline, but deployed widely to as many others in realtime feedback loops as much as possible in the ecosystem, not just inside the lab. All great progress in complex systems ultimately comes from successes and failures in the wild.

Today, Robocurve, an independent research firm, published its evaluations of Dream-1, the first model from the Dream Labs team. Against the major LLMs including Opus 5.5, Gemini 4 Argon, and GPT-6 Astra, the Dream-1 model is more performant on RoboDojo-RC Tier 1 (which runs 22 tests for evals on real world robotic arms), while being faster and cheaper than all of them:

That's just the beginning. Dream-1 handles tabletop tasks like picking, placing, and folding today, with wheeled mobile robots and dexterous hands next. The new funding goes primarily to compute the fuel these models run on, to expand the tasks and robots Dream-1 can handle and to support many more developers. It's a seed round sized for a frontier model lab with both a research mission and a global developer platform to build out.

A brain any robot can use. Learned from the way people move through the world, delivered through an API, and open for anyone to test. We believe this is how robotics finally goes general.

We're psyched to back Joel and the entire Dream Labs team as they build it.

Welcome, Dream Labs!

Try the Robot Playground and learn more at dreamlabs.ai.