Odyssey 开放世界模型 Odyssey-3 免费交互预览

Odyssey is making its world model Odyssey-3 publicly available. It generates interactive worlds in real time and achieves top scores on physics benchmarks, according to the company.

California-based AI company Odyssey has launched a public research preview of its world model Odyssey-3. Users can generate interactive environments from text prompts and explore them in real time. The model simulates physical processes and predicts how environments change based on specific actions. Developers can apply for API access.

Founders Oliver Cameron and Jeff Hawke first unveiled the model on September 15, with a focus on robotics, autonomous driving, and video games. What's new now is public access, more technical detail, and benchmark results.

Interactive AI worlds you can try right now

The free online demo runs on Odyssey-3 Flash. The model generates interactive environments from text descriptions. Users can choose between first-person and third-person perspectives, move through the generated world, trigger events, and watch the model respond in real time.

Odyssey-3 is built on an autoregressive diffusion transformer that continuously generates new video frames based on previous frames and user actions. According to Odyssey, the model learns physical relationships and cause-and-effect from visual observations during training.

Training data included internet videos with event descriptions, video game footage paired with the corresponding keyboard and mouse inputs, and simulated physical interactions. An extra training technique reduces the number of required compute steps, making real-time generation possible.

The base Odyssey-3 model has 14 billion parameters and generates video at 832 × 480 pixels, according to the official benchmark submission. Odyssey-3 Pro supports 1280 × 720 pixels.

Physics benchmark claims come with caveats

The more powerful Odyssey-3 Pro scores 66.1 points on the video-to-video benchmark from Physics-IQ Verified, according to the company. The test evaluates physical behavior across areas like fluid mechanics, optics, solid mechanics, magnetism, and thermodynamics. Models have to continue videos of real-world experiments, and their outputs are compared against actual outcomes.

That top score of 66.1 points comes with a big caveat, though. It's from a single test run where a selection method picked one of eight generated videos for each task. The benchmark rules require four test runs with standard deviation reported for any record claim. The reported record doesn't meet that bar. Without the selection method, Odyssey-3 Pro averaged 63.37 points across four runs. Both results appear on the official leaderboard but were submitted by Odyssey itself.

On the WorldMark benchmark, Odyssey-3 ranks first in three of four categories based on its own evaluation: First-Person Stylized (77.2), Third-Person Real (79.0), and Third-Person Stylized (76.3). In First-Person Real, it places third with 80.6 points. WorldMark tests how well models follow control instructions, how good the generated images look, and whether simulated worlds stay consistent over time.

One model for robots, drones, and GTA V

Odyssey wants to use the same world model across different tasks, from moving robotic arms to navigating drones to controlling game characters. For each application, the model gets paired with a specialized controller that translates its predictions into concrete commands.

In tests, an AI built on Odyssey-3 controlled multiple robotic arms, according to the company. A few dozen hours of demonstration data was enough for training. The robots could also recover from failed grasps on their own, even though those situations weren't part of the training data.

For humanoid robots, Odyssey is working with Swiss robotics company Flexion. Controllers that Flexion built on Odyssey-3 are said to perform more reliably than comparison models, even when conditions change.

Odyssey also built a drone controller trained on simulated flight data. In a virtual indoor environment, the drone could dodge obstacles and fly to specific targets on command, according to the company.

Odyssey-3 can play video games, too. The company showed an AI autonomously playing GTA V, steering vehicles, and fighting enemies. A controller trained on roughly two hours of GTA footage was also able to transfer its skills to Red Dead Redemption 2 without any extra training, moving a character on horseback.

Odyssey also plans to use its world model for training AI agents. In one demo, an AI agent received a task in natural language and tried to complete it through its own actions inside an Odyssey-3-generated environment. The goal is for agents to learn from the outcomes of their own actions, a potentially new training paradigm for AI.

The public research preview offers interactive environment generation for now, robotics and autonomous systems will need further work.

The race to build world models

Odyssey was founded in 2023 by Oliver Cameron and Jeff Hawke. In June 2026, the company raised $310 million from investors including Amazon and AMD Ventures.

World models aim to give AI systems an understanding of spatial and physical relationships. Google DeepMind is pursuing a similar approach with Genie 3 for generating interactive worlds. World Labs, the startup founded by Fei-Fei Li, is developing similar technology. AMD announced in late September that it plans to acquire World Labs for roughly $8.2 billion.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.