# Odyssey (AI lab)

> Source: https://aiwiki.ai/wiki/odyssey_ai
> Updated: 2026-09-16
> Fact-checked: 2026-09-16
> Categories: AI Companies, Artificial Intelligence, Research Organizations, World Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Odyssey (AI lab)." aiwiki.ai, 16 Sept 2026. https://aiwiki.ai/wiki/odyssey_ai
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Odyssey is an artificial intelligence lab that builds what it calls foundation world models: causal, action-conditioned systems trained to predict how the world evolves from one moment to the next, and to stream that prediction back as interactive video. It was founded in 2023 by Oliver Cameron and Jeff Hawke, two veterans of the self-driving industry, and runs offices in Palo Alto, London and Zurich [5]. In June 2026 the company raised a $310 million Series B at a $1.45 billion valuation led by Natural Capital [6], which TechCrunch reported brought its total funding to $337 million [7].

The company's public identity has shifted twice. It launched in 2024 as a visual-AI company aimed at film and games, pitching AI-generated "Hollywood-grade" long-form video built from real-world 3D capture [9]. Through 2025 and 2026 it repositioned around general-purpose [world models](https://aiwiki.ai/wiki/world_model), first as real-time interactive video and then as a control substrate for robots, vehicles, drones and game-playing agents [13][26]. Its web address moved with it: the site ran on odyssey.systems in 2024, used odyssey.ml and, for most of 2025, a landing page on odyssey.world, and was back on odyssey.systems by September 2026, with odyssey.ml subdomains such as the developer portal still in service [3][29].

## Founding and background

Cameron and Hawke both came out of autonomous driving. In the Series B announcement Cameron wrote that he, Hawke and the founding team had spent "more than a decade" working on teaching cars to navigate, and that the narrow road-facing world models built for that problem convinced them the same idea could generalize [6]. The Odyssey-3 post repeats the framing: the pair began working on autonomous vehicles and robotics in the 2010s, when general-purpose physical intelligence was a stated long-term ambition of the field, and founded Odyssey in 2023 on the belief that world models were a credible technical route to it [26].

Cameron was a vice president at Udacity when Voyage, a self-driving taxi startup, spun out of the company's autonomous-vehicle program in April 2017; Voyage planned to run a fleet built by retrofitting existing cars rather than designing vehicles from scratch [28]. Cameron was Voyage's co-founder and CEO [7][9]. [Cruise](https://aiwiki.ai/wiki/cruise) acquired Voyage in March 2021 [31], after which Cameron served as Cruise's VP of Product and, according to his investors, led the launch of the company's first fully driverless product in San Francisco before it scaled to three US cities [9][11]. Kyle Vogt, who founded Cruise, is among Odyssey's individual investors [6][7].

Hawke was a founding engineer at [Wayve](https://aiwiki.ai/wiki/wayve), the UK [autonomous driving](https://aiwiki.ai/wiki/autonomous_driving) company, where Air Street Capital credits him with pioneering the use of end-to-end models to drive on complex roads [9]. Odyssey lists him as Dr. Jeff Hawke, co-founder and CTO [5]. He writes much of the lab's technical commentary and presented Odyssey's research at RAAIS 2026 [4][18][20].

## Funding

| Announced | Round | Amount | Valuation | Lead | Other named participants |
| --- | --- | --- | --- | --- | --- |
| 8 July 2024 | Seed ("maiden round") | $8 million | not disclosed | not disclosed | Air Street Capital [9] |
| 13 November 2024 | Series A | $18 million | not disclosed | EQT Ventures | GV, Air Street Capital (repeat investors); brought total raised to $27 million [10][11] |
| 12 February 2026 | Strategic investment | not disclosed | not disclosed | not disclosed | NVentures (NVIDIA) and Samsung Next, announced jointly [8] |
| 17 June 2026 | Series B | $310 million | $1.45 billion | Natural Capital | Amazon, GV, AMD Ventures, EQT, IQT [6][7] |

TechCrunch put Odyssey's cumulative funding at $337 million after the Series B [7]. The disclosed rounds sum to $336 million; the $27 million post-Series A total reported by Maginative and TechCrunch is $1 million more than the $8 million seed plus the $18 million Series A, and no source explains the difference [7][9][10][11]. Writing about the May 2025 Odyssey-1 release, the same publication described the then-$27 million total as coming from EQT Ventures, [GV](https://aiwiki.ai/wiki/google_ventures) and Air Street Capital [11].

The February 2026 announcement named [NVentures](https://aiwiki.ai/wiki/nvidia_nventures), NVIDIA's venture arm, and Samsung Next as new investors and listed GV, EQT, Air Street Capital, Elad Gil, [Jeff Dean](https://aiwiki.ai/wiki/jeff_dean), Guillermo Rauch, Garry Tan and Kyle Vogt as existing backers [8]. Andy Duong, an investment director at Samsung Next, said the firm was "impressed with the rapid technical advances demonstrated by Odyssey-2 Pro, showing promising progress towards interactive world simulation, and the early signs of teaching artificial intelligence true cause-and-effect" [8]. The company did not disclose the size of that investment.

For the Series B, Odyssey named Jeff Dean, Elad Gil, Qasar Younis, Kyle Vogt and Garry Tan as existing investors joining the round, alongside unnamed researchers from OpenAI, DeepMind, MSL, Recursive and Thinking Machines [6]. Jay Zaveri, a general partner at Natural Capital, called Odyssey "our largest investment to date" and said the firm believes the company can "help define AI beyond language models" [6]. Odyssey's site separately lists Jeff Dean, Soumith Chintala, Max Jaderberg and Tim Rocktäschel among its backers, describing them respectively as chief scientist at DeepMind, CTO of Thinking Machines, president of Isomorphic Labs and a co-founder of Recursive [1].

Ed Catmull, a co-founder of Pixar and former president of Walt Disney Animation Studios, joined Odyssey's board of directors and invested in the company in December 2024 [12][11].

## Models and research

Odyssey publishes most model releases as blog posts. Two carry longer write-ups: PROWL-1 has an arXiv preprint [27], and Starchild-1 has a technical report credited to "Team Odyssey" and self-hosted at starchild.odyssey.ml rather than posted to a preprint server [21][30]. The lineage below follows the dates given on each post. Two of the names are retrospective: the May 2025 release was announced as an "interactive video model" and only retitled Odyssey-1 in early 2026, and the Odyssey-2 post was likewise reworded from "interactive AI video" to "world model" (see [Odyssey-2](https://aiwiki.ai/wiki/odyssey_2)) [29].

| Model or artifact | Announced | What it is |
| --- | --- | --- |
| Explorer | 18 December 2024 | Image-to-world model that turns a single image into a 3D world rendered as [Gaussian splats](https://aiwiki.ai/wiki/gaussian_splatting), aimed at film and gaming pipelines [12] |
| Odyssey-1 | 28 May 2025 | Research preview of a real-time playable world model, generating a frame every 40 ms and streaming at up to 30 FPS [13] |
| Odyssey-2 Alpha | not stated | Listed on Odyssey's research index between Odyssey-1 and Odyssey-2 as "a general world model with navigation input"; its entry links to a research video rather than a launch post [4] |
| [Odyssey-2](https://aiwiki.ai/wiki/odyssey_2) | 27 October 2025 | General-purpose world model that streams interactive video shaped by text prompts, producing a frame every 50 ms (20 FPS) [14] |
| Odyssey-2 Pro | 23 January 2026 | Larger Odyssey-2 model, the first served through the Odyssey developer API; streams 720p at 22 FPS [16] |
| Odyssey-2 Max | 21 April 2026 | Largest of the Odyssey-2 family, released in private beta; roughly 3x the parameters and 10x the training compute of Pro [19] |
| [PROWL-1](https://aiwiki.ai/wiki/prowl_1) | 12 May 2026 | Adversarial [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) framework in which an agent hunts for world-model failures and feeds them back as training data [20][27] |
| [Starchild-1](https://aiwiki.ai/wiki/starchild_1) | 17 May 2026 | World model that autoregressively generates synchronized audio and video in real time [21] |
| [Agora-1](https://aiwiki.ai/wiki/agora_1) | 18 May 2026 | Multi-agent world model letting up to four participants share one generated simulation, demonstrated on GoldenEye [22] |
| [CaliBench](https://aiwiki.ai/wiki/calibench) | 10 August 2026 | Benchmark testing whether world models reproduce the correct distribution of physical outcomes, not just plausible-looking single videos [25] |
| [Odyssey-3](https://aiwiki.ai/wiki/odyssey_3) | 15 September 2026 | Foundation world model used as the pretrained base for policies controlling robot arms, humanoids, cars, drones and video games; Odyssey says the backbone is kept frozen in the driving, drone and gaming experiments [26] |

Odyssey reports benchmark numbers for the Odyssey-2 family on VBench 2's physics sub-score and on the physics subset of PAI-Bench. It says Odyssey-2 Max raised the VBench 2 physics score from 49.67 to 58.52 and the PAI-Bench physics score from 91.67 to 93.02 relative to Odyssey-2 Pro, and that Odyssey-2 achieved 48.58 and 89.50 on the two measures [19]. The company's table also lists Cosmos-Predict2.5-14B, Cosmos-Predict2-14B, Cosmos-Predict2.5-2B and LingBot-World-Fast as comparison systems, and notes that some of the figures come from Odyssey's own measurements rather than the original benchmark papers [19]. Those comparisons have not, as published, been reproduced by an independent evaluator, and Odyssey's claim that Odyssey-2 Max "achieves the highest physics score among evaluated world models" is scoped to the models it chose to run [19].

## Technology and thesis

Odyssey's definition of a world model is narrow and mechanical: an action-conditioned dynamics model that, given the current state, an incoming action and a history of states and actions, predicts the next state as a video frame [13]. The company argues this is what separates a world model from a video model such as [Sora](https://aiwiki.ai/wiki/sora), [Veo](https://aiwiki.ai/wiki/veo_3), [Kling](https://aiwiki.ai/wiki/kling) or [Runway](https://aiwiki.ai/wiki/runway_ml): a bidirectional video model fixes past, present and future jointly from a prompt, which rules out real-time interaction because later frames would have to condition on actions the user has not taken yet [19].

The architecture Odyssey uses is an autoregressive [diffusion transformer](https://aiwiki.ai/wiki/diffusion_transformer) (AR DiT), which generates frames sequentially while conditioning each one on the frames and actions before it [17][19]. For Odyssey-2 Max the company described a proprietary KV cache enabling training on sequences up to 20 times longer than prior work with full backpropagation, causal attention combining local and global context, conditioning on latent-space embeddings so that arbitrary action formats can be handled, roofline-based inference-aware modelling, continuous flow matching in latent space instead of discrete tokenization, and distillation to a small number of denoising steps [19]. Training ran in three stages: large-scale video pretraining for general visual dynamics, then interaction and task conditioning, then a long-horizon stability phase to stop per-step errors accumulating during extended rollouts [19].

The lab's public argument for this direction is that language is too thin a record of reality to serve as a complete world model. Cameron's February 2026 essay makes the case that text omits body language, balance, weight and sound, and that video is the only data source large and rich enough to teach a machine how the world behaves [17]. The same essay argues that narrow world models already exist, pointing to the next-state prediction systems built by self-driving companies, and that generalizing them is the remaining work [17]. An earlier essay from December 2025 set out the companion argument: that next-state prediction works as a pretraining objective for the same reason next-token prediction does, and that a general learned simulator could displace the narrow, hand-built simulators each domain writes for itself, since a rigid-body physics engine is no use for simulating weather [15].

Hawke's February 2026 essay pushed back on the looseness of the term itself, proposing a four-way taxonomy of things being called world models [18]:

| Category | What it learns | Examples Hawke cites |
| --- | --- | --- |
| World models | How the world evolves | Odyssey, DeepMind's [Genie](https://aiwiki.ai/wiki/genie_3), LeCun's JEPA proposal |
| Spatial intelligence | How the world appears | [World Labs](https://aiwiki.ai/wiki/world_labs), 3D Gaussian splat reconstruction |
| Behaviour models | How to act within a world | General Intuition, Simile, Wayve's driver, [Physical Intelligence](https://aiwiki.ai/wiki/physical_intelligence) |
| Proxy world models | An abstraction of the world | Large language models with internal representations of simple world state |

He traces the canonical definition to Ha and Schmidhuber's 2018 work in a reinforcement learning context, and argues that spatial-intelligence systems, because they rely on explicit representations, have no clear path to general end-to-end learning [18].

Odyssey has framed its own progress against language-model history. In January 2026, launching Odyssey-2 Pro and its API, it called the moment "the GPT-2 moment for general-purpose world models" [16]. By the June 2026 Series B it said the field was approaching "the GPT-3 moment for world models" [6]. The Odyssey-3 post estimates that frontier world models remain roughly two orders of magnitude smaller than language models [26].

A fourth strand is the lab's interest in using world models as training environments. "Our Path to Superintelligence" argues that hand-built RL environments such as Minecraft and StarCraft impose a ceiling because they do not evolve with the agent, and that learned environments which update in response to interaction remove it [24]. PROWL-1 is the concrete version of this: an RL agent is rewarded for finding breakdowns in geometry, motion, visual consistency and action-conditioned dynamics, and the world model is fine-tuned on what the agent finds, creating an escalating curriculum [20][24]. The arXiv paper describes the mechanism as a KL-constrained adversarial curriculum with a prioritized adversarial trajectory buffer, initialized from a world model pretrained with [diffusion forcing](https://aiwiki.ai/wiki/diffusion_forcing) on human Minecraft data [20][27].

## Compute and partners

The Series B made Amazon Web Services Odyssey's preferred cloud provider, and Odyssey said it is working with Amazon's Annapurna Labs to optimize its world models for [AWS Trainium](https://aiwiki.ai/wiki/aws_trainium) [6]. Ron Diamant, a vice president and distinguished engineer at Amazon, said Trainium is "purpose-built for exactly this kind of scale" [6]. TechCrunch framed the Trainium work as positioning Amazon's silicon against NVIDIA's [7].

NVIDIA hardware nonetheless runs through the lab's history. The Odyssey-1 research preview streamed from clusters of H100 GPUs in the US and EU at a stated infrastructure cost of $1 to $2 per user-hour [13][11], and Odyssey-2 Max was trained on several hundred NVIDIA Blackwell B200 GPUs [19]. NVentures invested in February 2026 [8], and AMD Ventures joined the Series B [6][7].

With Odyssey-3 the company announced two application partnerships: a research collaboration with [Flexion](https://aiwiki.ai/wiki/flexion_robotics) on [humanoid](https://aiwiki.ai/wiki/humanoid_robot) control policies built on Odyssey-3 as a base model, and work with Poke & Wiggle on evaluating the model across different robot bodies, viewpoints and controls [26].

## People

Odyssey's published leadership page lists six people [5]:

| Name | Role |
| --- | --- |
| Oliver Cameron | Co-founder and CEO |
| Jeff Hawke | Co-founder and CTO |
| Fabian Güra | VP, Foundation Models (also shown as Distinguished Engineer on the same page) |
| James Grieve | VP, Applied Research |
| Jessica Inman | VP, Commercialization |
| Robert Grant | VP, Policy |

Ed Catmull sits on the board of directors [12]. Named authors on Odyssey's research writing include Jon Sadeghi (CaliBench) and Ahmet Hamdi Guzel (the multi-agent experience essay), in addition to Cameron and Hawke [3][23][25]. The PROWL preprint is credited to Ahmet H. Güzel, Jenny Seidenschwarz, Benjamin Graham, Jonathan Sadeghi, Jeffrey Hawke and Ilija Bogunovic [27].

The Odyssey-1 post named the full team behind that release, then fifteen technical and two operational staff [13]; by Odyssey-2 Max a year later the credits listed thirty-one people [19]. Odyssey says its team is recruited from DeepMind, Waymo, Tesla, Meta, Apple and xAI [2]; the February 2026 funding post gave the list as DeepMind, OpenAI, ByteDance, Tesla, Waymo, Meta, Wayve and Luma [8], and Air Street's 2024 post cited Cruise, Waymo, Wayve, Tesla, Microsoft, Meta and NVIDIA [9].

## Earlier product history

Odyssey's first public identity had little to do with robotics. The July 2024 seed announcement described a company training four separate models for geometry, materials, lighting and motion, feeding a data engine that combined real-world capture with 3D training data made by Hollywood VFX artists, with the goal of AI-generated long-form video that a director could control and version rather than re-prompt [9].

The November 2024 Series A funded the capture side of that plan. Odyssey built a 25-pound backpack rig carrying six cameras and two lidars, capable of 13.5K capture, and sent teams out to scan landscapes across California, with the stated intent of expanding to other states and countries [10]. Odyssey's own framing at the time was that generative models could not produce worlds that feel alive without training on large volumes of rich real-world 3D data [10]. TechCrunch later described the approach as a deliberate divergence from labs training on publicly available data alone [11].

Explorer, announced in December 2024 alongside Catmull's board appointment, was the model built on that data. It took an image and produced a 3D world represented as Gaussian splats, chosen because splat support was appearing in Unreal, Houdini, Blender, Maya, 3ds Max and After Effects, so generated worlds could be hand-edited in existing tools [12]. Odyssey said generations took about ten minutes on average at the time, and that it had trialled Explorer-generated worlds on the virtual production stage at Garden Studios in London [12].

The pivot came with Odyssey-1 in May 2025, which replaced offline 3D generation with real-time interactive video, and was completed by Odyssey-2 in October 2025, whose interface is a text box driving a live video stream [13][14]. Odyssey's research index still lists Explorer in its model lineage, described as "a purpose-built world model for film and gaming" [4]. The interactive-video product line itself continued: Odyssey-2 Pro shipped with three API endpoints (simulations, interactive streams and viewable streams) plus JavaScript and Python SDKs in January 2026 [16].

### Name and domains

Odyssey has published under the same name since its first funding announcement in July 2024 [9]. The website has not been as stable. Internet Archive captures show odyssey.systems serving the company's film-and-gaming site in 2024, including the Explorer launch material in December of that year [29]. By January 2025 the main site was on odyssey.ml, and through most of 2025 the company's public landing page lived on a third domain it also owned, odyssey.world: archived captures of odyssey.systems from March, April, August and October 2025, and of www.odyssey.ml from February 2025, redirect there, and odyssey.world served first an Odyssey "under construction" teaser and by September 2025 the interactive-video landing page, both declaring odyssey.systems as the canonical address, while blog posts and the demo stayed on odyssey.ml; from November 2025 odyssey.systems and odyssey.world redirected to odyssey.ml instead [29]. The Internet Archive still recorded odyssey.ml serving the full site on 4 September 2026; by 16 September 2026 odyssey.ml and www.odyssey.ml were returning permanent redirects to odyssey.systems, so the switch back happened inside that window [29]. The company's X handle, @odysseyml, and several service subdomains including the developer portal at developer.odyssey.ml and the Starchild-1 report host still sit on the older domain [3][26][30].

## Reception

Independent coverage has concentrated on Odyssey's fundraising. TechCrunch reported both the Series B and, earlier, the Odyssey-1 launch, and its hands-on assessment of that first public demo was blunt: the generated environments were "blurry and distorted, and unstable in the sense that their layouts don't always remain the same," with surroundings that could change if the user walked forward or turned around [11]. The same piece placed Odyssey among a crowded field chasing world models that includes DeepMind, World Labs, Microsoft and Decart, and noted that creative-industry reaction to the technology has been mixed, citing an Animation Guild study estimating large-scale disruption to US film and animation jobs [11].

Air Street Capital, which invested in the seed round, published its thesis in 2024 and compared the ambition to Pixar's transformation of film with CGI [9]. Nathan Benaich, who co-signed the post, disclosed there that he is also an angel investor in Wayve, where Hawke worked [9].

Hawke's taxonomy essay drew a public line between Odyssey's definition of a world model and neighbouring efforts, naming World Labs, General Intuition, Simile, Wayve and Physical Intelligence as belonging to other categories [18]. That framing is Odyssey's, not a settled consensus in the field.

Several of the lab's own superlatives should be read as company claims rather than established facts: "the world's first multimodal world model" for Starchild-1 [21], "materially advancing the state-of-the-art in physical accuracy" for Odyssey-2 Max and Odyssey-3 [19][2], and the GPT-2 and GPT-3 analogies for the field's maturity [16][6].

## See also

- [World model](https://aiwiki.ai/wiki/world_model)
- [Genie 3](https://aiwiki.ai/wiki/genie_3)
- [NVIDIA Cosmos](https://aiwiki.ai/wiki/nvidia_cosmos)
- [GAIA-4 (Wayve)](https://aiwiki.ai/wiki/gaia_4_wayve)
- [World Labs](https://aiwiki.ai/wiki/world_labs)
- [Physical AI](https://aiwiki.ai/wiki/physical_ai)
- [Robot foundation model](https://aiwiki.ai/wiki/robot_foundation_model)

## References

1. Odyssey, "About Odyssey." https://odyssey.systems/about
2. Odyssey, homepage. https://odyssey.systems/
3. Odyssey, "The latest from Odyssey" (blog index). https://odyssey.systems/writing
4. Odyssey, "A leading body of world model research." https://odyssey.systems/research
5. Odyssey, "Explore a Role with Odyssey" (careers). https://odyssey.systems/careers
6. Oliver Cameron, "Our $310 Million Fundraise to Accelerate World Simulation," Odyssey, June 17, 2026. https://odyssey.systems/our-series-b
7. Julie Bort, "World model maker Odyssey nabs $1.45B valuation backed by Amazon and other big names," TechCrunch, June 17, 2026. https://techcrunch.com/2026/06/17/world-model-maker-odyssey-nabs-1-45b-valuation-backed-by-amazon-and-other-big-names/
8. Oliver Cameron, "Odyssey Announces Investment from NVentures and Samsung Next," Odyssey, February 12, 2026. https://odyssey.systems/investment-from-nvidia-and-samsung
9. Air Street Press and Nathan Benaich, "Our investment in Odyssey," July 8, 2024. https://press.airstreet.com/p/our-investment-in-odyssey
10. Chris McKay, "Odyssey Raises $18M to Build AI Worlds Using Backpack-Mounted Cameras," Maginative, November 13, 2024. https://www.maginative.com/article/odyssey-raises-18m-to-build-ai-worlds-using-backpack-mounted-cameras/
11. Kyle Wiggers, "Odyssey's new AI model streams 3D interactive worlds," TechCrunch, May 28, 2025. https://techcrunch.com/2025/05/28/odysseys-new-ai-model-streams-3d-interactive-worlds/
12. Oliver Cameron, "World Models for Film, Gaming, and Beyond," Odyssey, December 18, 2024. https://odyssey.systems/introducing-explorer
13. Oliver Cameron, "Introducing Odyssey-1: A Playable World Model," Odyssey, May 28, 2025. https://odyssey.systems/introducing-odyssey-1
14. Oliver Cameron, "Introducing Odyssey-2: A General-Purpose World Model," Odyssey, October 27, 2025. https://odyssey.systems/introducing-odyssey-2
15. Oliver Cameron, "The Dawn of a World Simulator," Odyssey, December 20, 2025. https://odyssey.systems/the-dawn-of-a-world-simulator
16. Oliver Cameron, "The GPT-2 Moment for World Models Is Here," Odyssey, January 23, 2026. https://odyssey.systems/the-gpt-2-moment-for-world-models
17. Oliver Cameron, "Why We Must Build World Models," Odyssey, February 17, 2026. https://odyssey.systems/why-we-must-build-world-models
18. Jeff Hawke, "On the Origin of Species (of World Models)," Odyssey, February 27, 2026. https://odyssey.systems/on-the-origin-of-species-of-world-models
19. Oliver Cameron, "Introducing Odyssey-2 Max: Scaled World Simulation," Odyssey, April 21, 2026. https://odyssey.systems/introducing-odyssey-2-max
20. Jeff Hawke, "Introducing PROWL-1: Learning Through Discovery," Odyssey, May 12, 2026. https://odyssey.systems/introducing-prowl-1
21. Oliver Cameron, "Starchild-1: The First Real-Time Multimodal World Model," Odyssey, May 17, 2026. https://odyssey.systems/introducing-starchild-1
22. Oliver Cameron, "Agora-1: The Multi-Agent World Model," Odyssey, May 18, 2026. https://odyssey.systems/introducing-agora-1
23. Ahmet Hamdi Guzel, "The Era of Multi-Agent Imagined Experience," Odyssey, June 12, 2026. https://odyssey.systems/the-era-of-multi-agent-imagined-experience
24. Oliver Cameron, "Our Path to Superintelligence," Odyssey, July 8, 2026. https://odyssey.systems/our-path-to-superintelligence
25. Jon Sadeghi, "Introducing CaliBench: Are World Models Physically Calibrated?," Odyssey, August 10, 2026. https://odyssey.systems/introducing-calibench
26. Oliver Cameron and Jeff Hawke, "Introducing Odyssey-3: A General-Purpose Physical Intelligence," Odyssey, September 15, 2026. https://odyssey.systems/introducing-odyssey-3
27. Ahmet H. Güzel, Jenny Seidenschwarz, Benjamin Graham, Jonathan Sadeghi, Jeffrey Hawke and Ilija Bogunovic, "PROWL: Prioritized Regret-Driven Optimization for World Model Learning," arXiv:2605.18803. https://arxiv.org/abs/2605.18803
28. Darrell Etherington, "Udacity spin-out Voyage aims to build a full self-driving taxi business," TechCrunch, April 5, 2017. https://techcrunch.com/2017/04/05/udacity-spin-out-voyage-aims-to-build-a-full-self-driving-taxi-business/
29. Internet Archive Wayback Machine CDX index and snapshots for odyssey.ml and odyssey.systems. https://web.archive.org/web/20241218163434/https://odyssey.systems/
30. Team Odyssey, "Starchild-1" technical report (self-hosted PDF). https://starchild.odyssey.ml/starchild-1.pdf
31. Kirsten Korosec, "Cruise acquires self-driving startup Voyage", TechCrunch, 15 March 2021. https://techcrunch.com/2021/03/15/cruise-acquires-self-driving-startup-voyage/