# Dream Labs

> Source: https://aiwiki.ai/wiki/dream_labs
> Summary: Dream Labs (legally Dream Labs, Inc.) is a San Francisco artificial intelligence company that builds robot foundation models and sells access to them through a public API.
> Updated: 2026-10-27
> Fact-checked: 2026-10-27
> Categories: AI Companies, Embodied AI, Robotics Companies
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Dream Labs." aiwiki.ai, 27 Oct 2026. https://aiwiki.ai/wiki/dream_labs
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

*Not to be confused with Digital Dream Labs.*

**Dream Labs** (legally Dream Labs, Inc.) is a San Francisco artificial intelligence company that builds [robot foundation models](https://aiwiki.ai/wiki/robot_foundation_model) and sells access to them through a public API.[1][22] It was founded in 2026 by four former [NVIDIA](https://aiwiki.ai/wiki/nvidia) researchers who worked on NVIDIA's Project GR00T robotics research: Joel Jang (chief executive), Scott Reed, Yunhao Ge and You Liang Tan.[1][4][7] All four are authors of DreamZero, a February 2026 paper on robot policies that predict future video and robot actions together, which Dream Labs describes as the paper that introduced [world action models](https://aiwiki.ai/wiki/world_action_model) (WAMs).[14][3]

The company came out of stealth on October 19, 2026, announcing a $90 million seed round led by B Capital along with Kindred Ventures and launching its first model, [Dream-1](https://aiwiki.ai/wiki/dream_1), through the Dream API.[1][25] Dream Labs says Dream-1 learns from video of people doing everyday tasks and can control bimanual robots with parallel grippers that it was never trained on, with no data from those robots.[1][2] At launch the hosted model cost $3 per hour, or $2 per hour for a lighter Dream-1-flash variant, and the service was offered as an experimental research beta.[2][22]

## Overview

| Field | Value |
| --- | --- |
| Legal name | Dream Labs, Inc.[22] |
| Founded | 2026[1] |
| Headquarters | San Francisco, California[1] |
| Founders | Joel Jang, Scott Reed, Yunhao Ge, You Liang Tan[4] |
| Chief executive | Joel Jang[1] |
| Focus | Robot foundation models (world action models) offered through an API[1][2] |
| First model | Dream-1 and Dream-1-flash (October 2026)[2] |
| Funding | $90 million seed, announced October 19, 2026[1] |
| Valuation | Not disclosed |
| Website | dreamlabs.ai |

## History

### Research at NVIDIA

All four co-founders worked in NVIDIA's Generalist Embodied Agent Research (GEAR) group, which works on NVIDIA's Project GR00T robot foundation models, released as [Isaac GR00T](https://aiwiki.ai/wiki/isaac_gr00t).[10][11][12][13] Jang's own site described him as a senior research scientist at the GEAR lab who led the world-model team for Project GR00T; for [GR00T N1](https://aiwiki.ai/wiki/groot_n1), released in March 2025, he wrote that he led the video generation models and latent actions team.[10] Jang, Reed and Tan are all authors of the GR00T N1 paper.[18]

The group's papers trace a line from learning on unlabeled video to the world action model:

| Paper | Date | What it proposed | Dream Labs co-founders among authors |
| --- | --- | --- | --- |
| Latent Action Pretraining from Videos (LAPA)[17] | October 2024 | Pretraining a vision-language-action model on discrete "latent actions" learned from video, so that robot action labels are not needed during pretraining | Jang (equal-contribution first author with Seonghyeon Ye)[10] |
| GR00T N1[18] | March 2025 | An open foundation model for generalist humanoid robots | Jang, Reed, Tan |
| DreamGen[16] | May 2025 | Training robot policies on "neural trajectories," synthetic robot videos generated by video world models with pseudo-actions recovered afterwards | Jang (first author), Tan, Reed |
| DreamDojo[15] | February 2026 | A robot world model pretrained on 44,000 hours of egocentric human video | Tan, Jang |
| DreamZero ("World Action Models are Zero-shot Policies")[14] | February 2026 | A 14-billion-parameter world action model built on a pretrained video diffusion backbone | Ge, Tan, Reed, Jang |

The DreamZero paper reported more than a twofold improvement over state-of-the-art [vision-language-action models](https://aiwiki.ai/wiki/vision_language_action_model) in generalizing to new tasks and environments in real-robot experiments, real-time closed-loop control at 7 Hz, and adaptation to a new robot body with 30 minutes of play data while keeping zero-shot generalization.[14] Jang was one of the paper's four project leads, alongside first author Seonghyeon Ye, Yuke Zhu and [Jim Fan](https://aiwiki.ai/wiki/jim_fan); Ge and Tan were listed as core contributors.[12][10] Kindred Ventures, one of Dream Labs' investors, wrote that Jang coined the term "world action model."[4]

### Founding

Dream Labs' launch post says the 30-minute adaptation result was what convinced the team to start the company. After post-training DreamZero on 30 minutes of data from a YAM robot, the researchers saw it generalize zero-shot to tasks it had never seen, and the company wrote that this, together with falling robot hardware prices, made the timing feel right.[2] The post describes "six months of building" before the October 2026 launch.[2] The Korean outlet Money Today reported that Jang co-founded Dream Labs in April 2026.[8]

The company's team page describes "a small team in San Francisco" and shows a photo of twelve people; it lists Ai2, Georgia Tech, [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind), NVIDIA, Stanford University, Toyota Research Institute and UC Berkeley as places the team came from.[3] In October 2026 its careers page listed seven roles in machine learning, infrastructure, robot operations, security and product; five were based in San Francisco and two in San Francisco or Atlanta.[24]

### Launch and seed round

On October 19, 2026, Dream Labs announced the $90 million seed round and launched three products at once: the Dream API, a fine-tuning service and a browser-based Robot Playground.[1] Its launch post on X described Dream-1 as "a robot foundation model that controls bimanual robots it has never seen, on tasks it was never shown, in places it has never been, from $3/hour."[25] The company said it would spend the money mainly on compute, to widen the tasks and robots Dream-1 can handle and to support more users.[1]

## Founders

| Name | Role at Dream Labs | Background |
| --- | --- | --- |
| Joel Jang | Co-founder and CEO[1] | Senior research scientist at NVIDIA's GEAR lab, leading the world-model team for Project GR00T; B.S. in computer science from Korea University, M.S. in artificial intelligence from [KAIST](https://aiwiki.ai/wiki/kaist), and Ph.D. student at the University of Washington (listed as September 2023 to May 2025, on leave). His earlier research was on language models.[10][4] |
| Scott Reed | President and Chief Scientist[11] | Principal research scientist at NVIDIA and a founding member of the GEAR lab (2024-2026); before that, senior staff research scientist at DeepMind; Ph.D. from the University of Michigan (2016) under Honglak Lee.[11] First author of DeepMind's 2022 [Gato](https://aiwiki.ai/wiki/gato) paper, "A Generalist Agent."[19] |
| Yunhao Ge | Co-founder and CTO[12] | Research scientist at NVIDIA's GEAR group and Cosmos lab, working on DreamZero, Project GR00T and [NVIDIA Cosmos](https://aiwiki.ai/wiki/nvidia_cosmos); Ph.D. in computer science from the University of Southern California, and visiting Ph.D. student at Stanford.[12] |
| You Liang Tan | Co-founder and Chief Product Officer[4] | Senior research engineer at NVIDIA's GEAR group from September 2024, a founding core member of Project GR00T and core lead of the open-source effort for Isaac GR00T N1.x; earlier a research engineer in [Sergey Levine](https://aiwiki.ai/wiki/sergey_levine)'s RAIL lab at UC Berkeley, and a software engineer at Open Robotics who contributed to [ROS 2](https://aiwiki.ai/wiki/robot_operating_system) and Open-RMF; M.S. in computer science from Georgia Tech.[13] |

Reed's Gato used a single set of weights for text, images, games and robot control.[19][4] Tan is a co-author of the [Octo](https://aiwiki.ai/wiki/octo) generalist robot policy and of SERL, a software suite for sample-efficient robotic [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning), both from 2024.[20][21] Kindred Ventures' announcement wrote of Ge that he "led GR00T's video-model work" at NVIDIA.[4]

## Products

| Product | Description |
| --- | --- |
| Dream-1 | The company's first world action model available through a public API; returns robot actions from camera images, the robot's state and a task instruction[2][26] |
| Dream-1-flash | A faster, lighter version of Dream-1[2] |
| Dream API | Hosted inference for Dream-1; developers connect a robot with a Python client, and the company runs its models in data centers around the world[1][2] |
| Fine-tuning as a Service (FaaS) | Customers upload demonstrations from their own robot or task and receive a fine-tuned model ID, which the company says takes hours[1][26] |
| Robot Playground | A web page where anyone can type commands to a real robot in the San Francisco office and watch it live[1] |
| SDK | Client software licensed under Apache 2.0; it contains no model weights and runs no inference locally[22] |

### How Dream-1 works

Dream-1 jointly predicts video and actions from three cameras, one on the robot's head and one on each wrist.[2] For robots it supports natively it can output joint angles, but it can also predict end-effector poses (where each gripper should be and whether it is open or closed). Every two-armed robot with parallel grippers shares that action space, and the robot's own inverse-kinematics solver turns the poses into joint commands.[2] The company's explanation is that video is the one representation all robots share: if the model can predict what a robot will look like as it finishes a task, it can produce the actions that get it there.[2] The launch post contrasts this with earlier cross-embodiment work, the [Universal Manipulation Interface](https://aiwiki.ai/wiki/universal_manipulation_interface), which transferred policies across arms from wrist cameras alone but only when the arms used the same gripper.[2]

Dream-1 was trained on the YAM robot. The launch post reports zero-shot results in simulation on 27 robots it had never seen, from many manufacturers, among them the [Unitree](https://aiwiki.ai/wiki/unitree) R1, the [ROBOTIS](https://aiwiki.ai/wiki/robotis) AI Worker, Hugging Face's [SO-101](https://aiwiki.ai/wiki/so_101) and Willow Garage's PR2. Its real-robot clips use arms such as Anvil Robotics' OpenYAM, MakerMods' Maker Arm and Metal Arm, and Seeed Studio's reBot.[2] The company lists its own limitations: performance depends on the quality of each robot's inverse-kinematics solver, the model does not control the path the arm takes to reach a pose, and robots whose kinematics differ sharply from what it has seen may get infeasible predictions.[2] When zero-shot control is not good enough, the FaaS service fine-tunes the model on the customer's data. The launch post shows five minutes of data for two partner arms, 77 minutes for a Dexmate robot with Sharpa dexterous hands, and 98 minutes for a YAM fitted with different end effectors.[2] Fuller detail on the model and its evaluations is in the Dream-1 article.

### Scope at launch

The press release says that out of the box Dream-1 "can control any robot with grippers, zero-shot," but the same release states the no-extra-data baseline more narrowly, as two-armed robots with parallel grippers.[1] The technical launch post calls Dream-1 "a manipulation policy for bimanual robots with parallel grippers."[2] At launch it handled tabletop tasks such as picking and placing objects and folding laundry; the company said support for wheeled mobile robots and dexterous hands would follow.[1]

### Pricing and terms

Dream Labs' launch post priced Dream-1 at $3 per hour and Dream-1-flash at $2 per hour. Each call returns 1.6 seconds of actions, or 2,250 calls per hour.[2] An archived copy of the post from launch day gives the same figures.[2] Some coverage reported a different number: SiliconANGLE wrote that the lightweight version cost 75 cents per hour.[5]

The API Terms of Service (version 1.0, effective October 19, 2026) call the service "an experimental beta research preview" and license it for research, evaluation and development only. Production or commercial deployment, and any use in which a person must be inside the robot's hazard zone while it operates, require a separate written agreement.[22] Under the terms, Dream Labs may use customers' camera, video and robot data, instructions and model outputs to train its models unless a separate written agreement says otherwise.[22] The launch post warns that Dream-1 "will make mistakes" and has not been validated for safety-critical use.[2]

## Funding

| Round | Announced | Amount | Led by | Other named participants | Valuation |
| --- | --- | --- | --- | --- | --- |
| Seed | October 19, 2026 | $90 million | B Capital, along with Kindred Ventures[1] | A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC and Adrian Aoun[1]; Thomas Wolf of [Hugging Face](https://aiwiki.ai/wiki/hugging_face)[4]; Base Ventures[8] | Not disclosed |

Accounts of who led the round differ slightly. The Business Wire release says the round was "led by B Capital along with Kindred Ventures";[1] Kindred's own post says B Capital led and Kindred invested alongside it;[4] SiliconANGLE described the round as jointly led, and The Robot Report and The AI Insider wrote that B Capital and Kindred Ventures led it.[5][6][7] Kindred said its founder, Steve Jang, would join Dream Labs' board.[4] Money Today reported that Base Ventures was the only Korean-affiliated firm in the round.[8] Neither the company nor the outlets cited here reported a valuation.

B Capital general partner Daisy Cai said in the announcement that "a capable robot arm has gone from the price of a luxury car to the price of a high-end laptop," and that the team "did some of the defining work in robot learning at NVIDIA."[1] Kindred called it "a seed round sized for a frontier model lab."[4]

## Partners and early users

Dream Labs says it works with more than 20 hardware partners, whose supported robots cost between about $2,000 and $10,000, and that it tunes the model further for each partner's robot.[1] Money Today reported that the partners include Unitree and Robotis.[8] The launch post showed unseen-task rollouts from beta testers at 12 sites, among them Hugging Face, [Scale AI](https://aiwiki.ai/wiki/scale_ai), [AgiBot](https://aiwiki.ai/wiki/agibot), KAIST and Robocurve.[2]

Ryan Chan, co-founder and CEO of MakerMods, which sells $2,000 bimanual arms, was quoted in the launch release saying that with Dream-1 "our robot was up and running in a few minutes and attempting tasks we never programmed."[1]

## Evaluation and reception

Robocurve, which Dream Labs' launch post describes as a public benefit corporation that measures frontier robotics capabilities, evaluated Dream-1 before launch on its RoboDojo-RC Tier 1 benchmark of real-robot tasks.[2] As of October 27, 2026, Robocurve's leaderboard gave Dream-1 a mean score of 59.3% over 120 attempts at an estimated $0.15 per attempt with hosted inference. The next-highest configuration listed was [Gemini 4 Argon](https://aiwiki.ai/wiki/gemini_4_argon) (high, priority) at 48.2% and $2.34 per attempt.[23] Robocurve is also one of the beta-tester sites named in Dream Labs' launch post.[2]

Yicai's coverage called the result preliminary support for low-cost robot-model services. It added that benchmark scores are not the same as task success rates in real factories or homes, and do not show that the model is fit for long unsupervised operation. The outlet noted that Dream-1 covered only bimanual parallel-gripper manipulation, not mobile, legged, dexterous-hand or whole-body robots. It named NVIDIA's GR00T, [Physical Intelligence](https://aiwiki.ai/wiki/physical_intelligence)'s pi models and [Gemini Robotics](https://aiwiki.ai/wiki/gemini_robotics) as competing robot foundation model efforts, and said Dream Labs' difference lay in how it sold the technology: as an hourly API with fine-tuning for new hardware.[9] Kindred's announcement listed [Ant Group](https://aiwiki.ai/wiki/ant_group), [Rhoda AI](https://aiwiki.ai/wiki/rhoda_ai), Sereact and [Mimic Robotics](https://aiwiki.ai/wiki/mimic_robotics) among other organizations working on world action models.[4]

## See also

- [Dream-1](https://aiwiki.ai/wiki/dream_1)
- [World action model](https://aiwiki.ai/wiki/world_action_model)
- [Robot foundation model](https://aiwiki.ai/wiki/robot_foundation_model)
- [Isaac GR00T](https://aiwiki.ai/wiki/isaac_gr00t)
- [Vision-language-action model](https://aiwiki.ai/wiki/vision_language_action_model)
- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)

## References

1. Business Wire. ["Dream Labs Raises $90M Seed to Build the Brain for Everyday Robots"](https://www.morningstar.com/news/business-wire/20261019230613/dream-labs-raises-90m-seed-to-build-the-brain-for-everyday-robots). Press release, October 19, 2026 (syndicated copy on Morningstar).
2. Dream Labs. ["Zero-Shot to New Robots: Introducing Dream-1"](https://dreamlabs.ai/research/dream-1/). October 2026. Launch-day copy: [Internet Archive, October 19, 2026](https://web.archive.org/web/20261019190211/https://dreamlabs.ai/research/dream-1/).
3. Dream Labs. ["Our team"](https://dreamlabs.ai/team/). Accessed October 27, 2026.
4. Steve Jang and Shryas Bhurat. ["Dream Labs: Building the Brain for Everyday Robots"](https://kindredventures.com/announcement/dream-labs-building-the-brain-for-everyday-robots/). Kindred Ventures, October 19, 2026.
5. Mike Wheatley. ["Dream Labs raises $90M to help build more intelligent robots with simple video feeds"](https://siliconangle.com/2026/10/19/dream-labs-raises-90m-to-help-build-more-intelligent-robots-with-simple-video-feeds/). SiliconANGLE, October 19, 2026.
6. The Robot Report Staff. ["Dream Labs raises $90M and launches Dream API"](https://www.therobotreport.com/dream-labs-raises-90m-and-launches-dream-api/). The Robot Report, October 22, 2026.
7. Greg Bock. ["Dream Labs Raises $90M in Seed Funding to Build Foundation Models for Any Robot"](https://theaiinsider.tech/2026/10/19/dream-labs-raises-90m-in-seed-funding-to-build-foundation-models-for-any-robot/). The AI Insider, October 19, 2026.
8. Choi Tae-bum. ["Korean American CEO in his 20s raises $120 billion for AI-powered robot brains, drawing US investors in droves"](https://www.mt.co.kr/index.php/en/future/2026/10/20/2026102009494495865). Money Today (Unicorn Factory), English edition, October 20, 2026. The headline's figure refers to 120 billion won ($90 million).
9. Lü Qian. [机器人行业"鸡生蛋"难题，被前英伟达团队撕开一道口子](https://www.yicai.com/news/103400978.html) (The robotics industry's "chicken-and-egg" problem, cracked open by a former NVIDIA team). Yicai, October 20, 2026. In Chinese.
10. Joel Jang. ["Joel Jang"](https://joeljang.github.io/). Personal website. Accessed October 27, 2026.
11. Scott Reed. ["Scott Reed"](https://scottreed.ai/). Personal website. Accessed October 27, 2026.
12. Yunhao Ge. ["Yunhao Ge"](https://gyhandy.github.io/). Personal website. Accessed October 27, 2026.
13. You Liang Tan. ["You Liang's Personal Site"](https://youliangtan.github.io/). Personal website. Accessed October 27, 2026.
14. Seonghyeon Ye, Yunhao Ge, et al. ["World Action Models are Zero-shot Policies"](https://arxiv.org/abs/2602.15922). arXiv:2602.15922, February 17, 2026.
15. Shenyuan Gao, William Liang, et al. ["DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos"](https://arxiv.org/abs/2602.06949). arXiv:2602.06949, February 6, 2026.
16. Joel Jang, Seonghyeon Ye, et al. ["DreamGen: Unlocking Generalization in Robot Learning through Video World Models"](https://arxiv.org/abs/2505.12705). arXiv:2505.12705, May 19, 2025.
17. Seonghyeon Ye, Joel Jang, et al. ["Latent Action Pretraining from Videos"](https://arxiv.org/abs/2410.11758). arXiv:2410.11758, October 15, 2024.
18. NVIDIA (Johan Bjorck et al.). ["GR00T N1: An Open Foundation Model for Generalist Humanoid Robots"](https://arxiv.org/abs/2503.14734). arXiv:2503.14734, March 18, 2025.
19. Scott Reed, Konrad Zolna, et al. ["A Generalist Agent"](https://arxiv.org/abs/2205.06175). arXiv:2205.06175, May 12, 2022.
20. Octo Model Team (Dibya Ghosh et al.). ["Octo: An Open-Source Generalist Robot Policy"](https://arxiv.org/abs/2405.12213). arXiv:2405.12213, May 20, 2024.
21. Jianlan Luo, Zheyuan Hu, Charles Xu, You Liang Tan, et al. ["SERL: A Software Suite for Sample-Efficient Robotic Reinforcement Learning"](https://arxiv.org/abs/2401.16013). arXiv:2401.16013, January 29, 2024.
22. Dream Labs, Inc. ["Dreamlabs API Terms of Service"](https://platform.dreamlabs.ai/legal/api-terms). Version 1.0, effective October 19, 2026. Sections 4.4, 4.7, 8.4, 10.7 and 14.2.
23. Robocurve. ["Real-world evaluations of physical AI"](https://robocurve.org/). RoboDojo-RC Tier 1 score and cost tables. Accessed October 27, 2026.
24. Dream Labs. ["Careers"](https://dreamlabs.ai/careers/). Accessed October 27, 2026.
25. Dream Labs (@dreamlabs_ai). [Launch post](https://x.com/dreamlabs_ai/status/2112230668805828801). X, October 19, 2026.
26. Dream Labs. ["Dream Labs"](https://dreamlabs.ai/). Official homepage. Accessed October 27, 2026.