# Embo (company)

> Source: https://aiwiki.ai/wiki/embo_ai
> Updated: 2026-10-27
> Fact-checked: 2026-10-27
> Categories: AI Companies, Embodied AI, Robotics Companies, World Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Embo (company)." aiwiki.ai, 27 Oct 2026. https://aiwiki.ai/wiki/embo_ai
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Embo** is a San Francisco artificial intelligence startup that develops [world models](https://aiwiki.ai/wiki/world_model) for robots.[1][2][4] It was co-founded by Danijar Hafner and Wilson Yan, two former research scientists at [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind).[2][3] Hafner created the [Dreamer](https://aiwiki.ai/wiki/dreamer) series of model-based [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) agents, and in 2025 he and Yan were equal-contribution authors of Dreamer 4, an agent trained inside a learned world model of the video game Minecraft.[4][5]

In a profile published in September 2026, MIT Technology Review described the startup as in stealth, with [humanoid robots](https://aiwiki.ai/wiki/humanoid_robot) in its office but no name on the door. In late October 2026 the embo.com homepage showed only a logo.[4][6] In March 2026 The Information reported, according to Techmeme's summary of the story, that Embo was in talks to raise a seed round of more than $100 million led by [Andreessen Horowitz](https://aiwiki.ai/wiki/andreessen_horowitz), with [Khosla Ventures](https://aiwiki.ai/wiki/khosla_ventures), DST Global and Striker participating.[1][2] The report described talks rather than a completed financing. Venture firm Sunflower Capital lists Embo among its portfolio companies without saying when it invested or how much.[7]

## Overview

| Field | Detail |
| --- | --- |
| Founded | 2025 (Hafner left Google DeepMind in the fall of 2025 to form the startup)[4] |
| Founders | Danijar Hafner, Wilson Yan[3] |
| Headquarters | San Francisco (SoMa district)[4] |
| Focus | World models for robotics[2][7] |
| Status | Stealth (September 2026 profile); logo-only website (October 2026)[4][6] |
| Website | embo.com[6][7] |
| Reported seed round | More than $100 million, reported as in talks (March 2026)[1][2] |
| Reported investors | Andreessen Horowitz (reported lead), Khosla, DST Global, Striker[2]; Sunflower Capital (portfolio listing)[7] |
| Valuation | Not disclosed[8] |

The company is unrelated to EMBO, the European Molecular Biology Organization.

## History

### Background and founding

Hafner became a student researcher at [Google Brain](https://aiwiki.ai/wiki/google_brain) in 2015, while he was a second-year engineering undergraduate at the Hasso Plattner Institute in Potsdam, and went on to hold a dozen internships and other positions at Google, according to MIT Technology Review.[4] He was previously a Staff Research Scientist at Google DeepMind in San Francisco.[9] On November 5, 2025, the Chinese technology site 36Kr, carrying a report by the outlet Xinzhiyuan, said that Hafner had written on X that it was his last day at DeepMind.[10] MIT Technology Review later reported that he had left Google DeepMind in the fall of 2025 to form his new startup.[4]

Yan studied computer science and applied mathematics at UC Berkeley and completed a PhD there advised by [Pieter Abbeel](https://aiwiki.ai/wiki/pieter_abbeel).[11][12] His resume lists a position as a senior research scientist in Google DeepMind's generative AI group starting in May 2024, "working on large-scale world models for decision making."[12] The two had worked together before Embo: Hafner was a co-author of Yan's 2022 TECO video-generation paper, and the pair shared first authorship of Dreamer 4.[5][13] Both were described as former Google DeepMind researchers when The Information reported on the company in March 2026.[1][3]

### The Information report (March 2026)

On March 4, 2026, Kevin McLaughlin of The Information reported that former Google DeepMind researchers were raising $100 million to build world models.[1] In a post on X linking the story, McLaughlin named the researchers as Hafner and Yan and the startup as Embo, said it was "developing world models for robotics," and called the financing "one of the largest seed rounds to date for a robotics startup."[3] Techmeme's summary of the report said Embo was in talks to raise a seed round of more than $100 million led by a16z, with Khosla, DST Global and Striker participating, and described Embo as "an AI startup co-founded by former Google DeepMind research scientists."[2]

### MIT Technology Review profile (September 2026)

On September 8, 2026, MIT Technology Review named Hafner, then 31, to its 2026 list of Innovators Under 35.[4] The profile, written by Mat Honan, described a visit earlier that year to Hafner's office in San Francisco's SoMa district. The startup was "still in stealth mode" and did not have its name on the door. Only one other person was there and there was little furniture, but humanoid robots "of various shapes and sizes" hung from racks running down the center of the room. According to the profile, Hafner imports the humanoids from China.[4]

Hafner would not say much about the venture. He described it as a continuation of his long-running work on enabling AI to handle environments it has not encountered in training, and said only that he "was interested in solving a problem" that "would change the world." The profile did not name the company.[4]

### Website and investor listing

Sunflower Capital's portfolio page lists Embo with the description "World models for robotics" and links it to the domain embo.com.[7] Sunflower describes itself as a firm that "partners at the earliest stage with companies building foundations for modern enterprises, critical industries, and the physical world."[7] In late October 2026 the embo.com homepage consisted of a white logo on a black background, with no text, product information or job listings.[6]

## Technology and approach

Public descriptions of Embo's technology come from press coverage and from the founders' earlier research. The Information and Techmeme described the company as building world models for robotics.[1][2] MIT Technology Review explained Hafner's method as model-based reinforcement learning. He builds world models, AI models designed to emulate physical reality, and trains agents inside them. The agent treats the world model as a simulation of the real world, learns how to act in it, and uses that experience to predict future outcomes. The magazine said this allows agents, and the robots they control, to handle unfamiliar situations in the real world.[4]

The profile gave the example of a robot sent into a person's home, which must cope with a floor plan and furniture it has never seen. It said that Hafner's technique lets agents and robots carry out complicated tasks without the real-world trial-and-error training traditionally used in robotics.[4] Hafner's X profile describes his work as "Building general agents for the physical world."[14]

The founders' most recent paper before the company, "Training Agents Inside of Scalable World Models" (Dreamer 4), was posted to arXiv on September 29, 2025. Hafner and Yan were listed as equal contributors and Timothy Lillicrap as third author, all at Google DeepMind in San Francisco.[5] Dreamer 4 learns control tasks by reinforcement learning inside a world model that runs real-time interactive inference on a single GPU, using a "shortcut forcing" objective and an efficient transformer architecture. The world model takes most of its knowledge from unlabeled videos and needs only a small amount of action-labeled data. The paper reported that Dreamer 4 was the first agent to obtain diamonds in Minecraft purely from offline data, without interacting with the game. The task requires choosing sequences of more than 20,000 mouse and keyboard actions from raw pixels. The authors also reported a large improvement over OpenAI's VPT offline agent while using 100 times less data.[5]

The authors tied the offline setting to "practical applications such as robotics where learning from environment interaction can be unsafe and slow." They contrasted their approach with controllable video models such as [Genie 3](https://aiwiki.ai/wiki/genie_3), which they said still struggle to learn the precise physics of object interactions and often need many GPUs to simulate a single scene in real time.[5]

## Research background

Hafner's research has centered on learned world models for planning and control,[9] and Yan's on scalable video generation models.[12] The table lists selected earlier papers by the founders.

| Year | Work | Founder authors | Description |
| --- | --- | --- | --- |
| 2018 | PlaNet ("Learning Latent Dynamics for Planning from Pixels") | Hafner (first author) | A purely model-based agent that learns environment dynamics from images and chooses actions by fast online planning in latent space[15] |
| 2019 | Dreamer ("Dream to Control: Learning Behaviors by Latent Imagination") | Hafner (first author) | Learns long-horizon behaviors from images purely by latent imagination; exceeded earlier approaches on 20 visual control tasks[16] |
| 2020 | DreamerV2 ("Mastering Atari with Discrete World Models") | Hafner (first author) | First agent to reach human-level performance on the 55-task Atari benchmark by learning behaviors inside a separately trained world model[17] |
| 2021 | VideoGPT | Yan (first author) | Video generation with a VQ-VAE and a GPT-style transformer over discrete latent codes[18] |
| 2022 | DayDreamer ("DayDreamer: World Models for Physical Robot Learning") | Hafner | Applied Dreamer to four physical robots without simulators; a quadruped learned to roll off its back, stand up and walk in one hour[19] |
| 2022 | TECO ("Temporally Consistent Transformers for Video Generation") | Yan (first author), Hafner | A video generation model aimed at long-term temporal consistency; published at ICML 2023[13][11] |
| 2023 | DreamerV3 ("Mastering Diverse Domains through World Models"; in Nature, April 2025) | Hafner (first author) | A single configuration that outperforms specialized methods across more than 150 tasks; described by the authors as the first algorithm to collect diamonds in Minecraft from scratch without human data or curricula[20] |
| 2024 | Large World Model ("World Model on Million-Length Video And Language With Blockwise RingAttention") | Yan | Open-source 7B-parameter models that process text documents and videos of more than 1 million tokens[21] |
| 2025 | Dreamer 4 ("Training Agents Inside of Scalable World Models") | Hafner and Yan (equal contribution) | First agent to obtain diamonds in Minecraft purely from offline data[5] |

DayDreamer applied Dreamer to physical robots. Besides the quadruped, the paper reported that Dreamer learned pick-and-place on two robotic arms directly from camera images and sparse rewards, and that the quadruped adapted within 10 minutes to withstand being pushed.[19] The Large World Model paper, which Yan wrote with Hao Liu, Matei Zaharia and Pieter Abbeel at Berkeley, used [Ring Attention](https://aiwiki.ai/wiki/ring_attention) and extended context length progressively from 4K to 1 million tokens.[21]

## Founders

### Danijar Hafner

Hafner grew up in a rural town in northeastern Germany, where his parents were both classical musicians, and learned programming from a neighbor, according to MIT Technology Review.[4] His website says he received his PhD at the University of Toronto with [Jimmy Ba](https://aiwiki.ai/wiki/jimmy_ba), was a visiting student at UC Berkeley with Pieter Abbeel, and completed an MRes at University College London and the Gatsby Unit with Timothy Lillicrap and Karl Friston. It also says he interned at Google Brain for six years and was named a Vanier Scholar. His PhD thesis is titled "Embodied Intelligence Through World Models."[9]

At Google he worked with [Geoffrey Hinton](https://aiwiki.ai/wiki/geoffrey_hinton) and [Ashish Vaswani](https://aiwiki.ai/wiki/ashish_vaswani), according to MIT Technology Review.[4] Lillicrap, a former manager and co-author, told the magazine: "I get to interact with a lot of really smart people in research at Google, and he easily sits in the top half of 1%." He added: "In many cases he would build, single-handedly, things it would take entire teams of engineers to build."[4]

### Wilson Yan

According to his resume, Yan earned a bachelor's degree in computer science and applied mathematics at UC Berkeley from 2016 to 2020. During that time he was a research assistant in the Robot Learning Lab of the Berkeley Artificial Intelligence Research lab from 2018 to 2020. He then began a PhD at Berkeley under Pieter Abbeel, during which he interned at [Perplexity](https://aiwiki.ai/wiki/perplexity) (October 2022 to March 2023) and in Meta's generative AI group (June to October 2023).[12] His homepage says he completed the PhD and was previously a senior research scientist at Google DeepMind, and describes his current research as "focused on world models for control."[11] His papers include VideoGPT, TECO, the Large World Model, ElasticTok (a variable-length tokenizer for images and long videos, ICLR 2025) and Dreamer 4.[11]

## Funding

| Date | Event | Amount | Investors | Status |
| --- | --- | --- | --- | --- |
| March 4, 2026 | Seed round reported by The Information | More than $100 million | Andreessen Horowitz (lead), Khosla, DST Global, Striker | Reported as in talks[1][2] |
| Not stated | Sunflower Capital portfolio listing | Not disclosed | Sunflower Capital | Listed as a portfolio company (October 2026)[7] |

No valuation was reported with the March 2026 talks.[8] Debarghya Das's "The Ultimate Neolab List" (updated October 24, 2026) includes Embo and gives it a "Clean" call, the first of the list's four categories (Clean, Mostly, Borderline and Not a neolab). Das summarizes the reasoning as "Elite researchers, pre-revenue, compute-heavy world models, software first, $100M+ seed."[8] The list shows a valuation of about $570 million for Embo but marks it as imputed and not from a closed round. Das's note says the figure was calculated by assuming that the reported $100 million bought about 17.5 percent of the company. It is not a reported valuation.[8]

## Context

An April 2026 AWS Startups article by Dhruv Sharma of AWS and Nikhil Suresh, an investor at Striker Venture Partners, on world models in robotics noted that many of the people behind companies in the field are researchers rather than product managers. It gave as examples that "Hafner and Yan moved from the Dreamer series to Embo," that Karol Hausman left DeepMind's robotics team to co-found [Physical Intelligence](https://aiwiki.ai/wiki/physical_intelligence), and that Yann LeCun moved from Meta AI to [AMI Labs](https://aiwiki.ai/wiki/ami_labs). The authors wrote that the willingness of such researchers to leave corporate research labs to commercialize their science "signals a major change."[22]

## See also

- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)
- [Dreamer (reinforcement learning)](https://aiwiki.ai/wiki/dreamer)
- [World model](https://aiwiki.ai/wiki/world_model)
- [Embodied AI](https://aiwiki.ai/wiki/embodied_ai)
- [Robot learning](https://aiwiki.ai/wiki/robot_learning)
- [General Intuition](https://aiwiki.ai/wiki/general_intuition)
- [World Labs](https://aiwiki.ai/wiki/world_labs)

## References

1. Kevin McLaughlin, ["Exclusive: Ex-Google DeepMind Researchers Raising $100 Million to Build 'World Models'"](https://www.theinformation.com/briefings/exclusive-ex-google-deepmind-researchers-raising-100-million-build-world-models), The Information, March 4, 2026 (subscription required).
2. Techmeme, ["Source: Embo, which is developing world models for robotics, is in talks to raise a $100M+ seed led by a16z, with Khosla, DST Global, and Striker participating (Kevin McLaughlin/The Information)"](https://www.techmeme.com/260304/p37), March 4, 2026.
3. Kevin McLaughlin (@KevKubernetes), ["Exclusive: Ex-Google DeepMind researchers Danijar Hafner and Wilson Yan are raising $100 million+ for their new startup Embo"](https://x.com/KevKubernetes/status/2029306254963949702), post on X, March 4, 2026.
4. Mat Honan, ["This entrepreneur is developing agents that can plan ahead"](https://www.technologyreview.com/2026/09/08/1142088/danijar-hafner-developing-plan-ahead-agents/), MIT Technology Review, September 8, 2026.
5. Danijar Hafner, Wilson Yan and Timothy Lillicrap, ["Training Agents Inside of Scalable World Models"](https://arxiv.org/abs/2509.24527), arXiv:2509.24527, September 29, 2025.
6. Embo, ["Embo"](https://embo.com/), company website, accessed October 27, 2026.
7. Sunflower Capital, ["Our Portfolio"](https://sunflowercapital.co/portfolio), accessed October 27, 2026.
8. Debarghya Das, ["The Ultimate Neolab List"](https://debarghyadas.com/neolabs/), updated October 24, 2026, accessed October 27, 2026.
9. Danijar Hafner, ["Danijar Hafner"](https://danijar.com/), personal website, accessed October 27, 2026.
10. Xinzhiyuan, ["Google's Dreamer Guru Exits Company, Reveals Missed Out on Transformer"](https://eu.36kr.com/en/p/3539389679104131), 36Kr, November 5, 2025.
11. Wilson Yan, ["Wilson Yan"](https://wilsonyan.com/), personal website, accessed October 27, 2026.
12. Wilson Yan, ["Resume"](https://wilsonyan.com/data/WilsonYanResume.pdf), PDF linked from personal website, accessed October 27, 2026.
13. Wilson Yan, Danijar Hafner, Stephen James and Pieter Abbeel, ["Temporally Consistent Transformers for Video Generation"](https://arxiv.org/abs/2210.02396), arXiv:2210.02396, October 5, 2022.
14. Danijar Hafner (@danijarh), [profile on X](https://x.com/danijarh), accessed October 27, 2026.
15. Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee and James Davidson, ["Learning Latent Dynamics for Planning from Pixels"](https://arxiv.org/abs/1811.04551), arXiv:1811.04551, November 12, 2018.
16. Danijar Hafner, Timothy Lillicrap, Jimmy Ba and Mohammad Norouzi, ["Dream to Control: Learning Behaviors by Latent Imagination"](https://arxiv.org/abs/1912.01603), arXiv:1912.01603, December 3, 2019.
17. Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi and Jimmy Ba, ["Mastering Atari with Discrete World Models"](https://arxiv.org/abs/2010.02193), arXiv:2010.02193, October 5, 2020.
18. Wilson Yan, Yunzhi Zhang, Pieter Abbeel and Aravind Srinivas, ["VideoGPT: Video Generation using VQ-VAE and Transformers"](https://arxiv.org/abs/2104.10157), arXiv:2104.10157, April 20, 2021.
19. Philipp Wu, Alejandro Escontrela, Danijar Hafner, Ken Goldberg and Pieter Abbeel, ["DayDreamer: World Models for Physical Robot Learning"](https://arxiv.org/abs/2206.14176), arXiv:2206.14176, June 28, 2022.
20. Danijar Hafner, Jurgis Pasukonis, Jimmy Ba and Timothy Lillicrap, ["Mastering diverse control tasks through world models"](https://www.nature.com/articles/s41586-025-08744-2), Nature 640, 647-653, published online April 2, 2025.
21. Hao Liu, Wilson Yan, Matei Zaharia and Pieter Abbeel, ["World Model on Million-Length Video And Language With Blockwise RingAttention"](https://arxiv.org/abs/2402.08268), arXiv:2402.08268, February 13, 2024.
22. Dhruv Sharma and Nikhil Suresh, ["World Models and the multimodal AI stack: rewriting the rules of robotics"](https://aws.amazon.com/aws-startups/learn/world-models-and-the-multimodal-ai-stack-rewriting-the-rules-of-robotics/), AWS Startups, April 28, 2026.