Odyssey launches Odyssey-3: one world model drives robots, cars, and drones
Odyssey, founded by self-driving car veterans, has released Odyssey-3, a model that predicts how visual scenes unfold. Based on visual data, it controls robots, cars, and drones, generates training environments, and plays video games. Record results: 66.1 points on Physics-IQ Verified (ahead of Nvidia and Black Forest Labs), first place in three WorldMark categories. Investors include Amazon and AMD Ventures, at a $1.45 billion valuation. • Flexion — humanoid robots controlled by Odyssey-3 adapt to lighting changes • Driving control on real roads in India using just 20 hours of training data • Drones fly indoors; the model generates environments for AI agents and plays GTA V
AI-processed from TNW; edited by Hamidun News
Odyssey, a Palo Alto-based company, has launched Odyssey-3, a world model capable of controlling robotic arms, humanoid robots, cars, and drones, as well as generating training environments for AI and playing video games. The research preview is already available.
World models are trained on massive amounts of visual data, learning to predict how scenes evolve and how actions affect what happens next. Odyssey claims its model gives machines a general understanding of the world before they're specialized for specific tasks. "Odyssey-3 is a major step toward a unified intelligence that can understand and act in the world around us," said Oliver Cameron, co-founder and CEO of Odyssey.
Benchmark leadership
Odyssey-3 Pro scored 66.1 points on the video-to-video Physics-IQ Verified test — the highest published result on that leaderboard. The benchmark, developed by Anates Labs and DeepMind, asks models to continue videos of real physical experiments; the result reflects the best of eight attempts. Models from Nvidia and Black Forest Labs are also part of the comparison.
The model also took first place in three of the four categories on the WorldMark benchmark (according to Odyssey's own evaluation).
Six areas of application
Odyssey demonstrated the model's performance across the following domains:
Robotic arms. With dozens of hours of demonstrations, the model carried out manipulation tasks and recovered from errors that weren't present in the training data — for example, a missed object grasp.
Humanoid robots. Flexion, a company specializing in humanoids, built control policies on top of Odyssey-3 that handled lighting changes that caused baseline models to fail.
Driving control. Odyssey tested the model's ability to drive a car on real roads in India; the policy was trained on roughly 20 hours of data.
Drone control. Odyssey-3 controlled drones flying indoors.
Environment synthesis. The model generated virtual environments for training AI agents.
Video games. Odyssey-3 successfully played Grand Theft Auto V.
Funding and the competitive landscape
Odyssey was founded by veterans of the self-driving car industry. In June, the company raised funding at a $1.45 billion valuation; investors include Amazon and AMD Ventures.
The company operates in a rapidly growing world models sector. AMD recently agreed to acquire Fei-Fei Li's company for $8.2 billion in stock. Another company, founded by former DeepMind researchers who worked on world models, was in talks to raise $700 million. Nvidia offers its own open-source world model. As a result, Odyssey finds itself competing against formidable players: Nvidia (already embedded in the ecosystem), AMD (through its acquisition), and well-funded startups.
- Oct 27, 2026Anthropic Files S-1 with the SEC: IPO Under Ticker ANT and $16.3 Billion in Six Months
- Oct 27, 2026Qwen3.8 Tried to Upload a File to Alibaba's Cloud in Beijing From Inside Cohere's Sandbox
- Oct 26, 2026Anthropic files S-1 with SEC, plans Nasdaq listing under ticker ANT
- Oct 26, 2026Claude Opus 5.5 wins Code2Video-AA with 1536 Elo, beating GPT-6 Astra by 109
Key Questions Answered
What are the key details on Benchmark leadership?
Odyssey-3 Pro scored 66.1 points on the video-to-video Physics-IQ Verified test — the highest published result on that leaderboard. The benchmark, developed by Anates Labs and DeepMind, asks models to continue videos of real physical experiments; the result reflects the best of eight attempts. Models from Nvidia and Black Forest Labs are also part of the comparison. The model also took first place in three of the four categories on the WorldMark benchmark (according to Odyssey's own evaluation).
What are the key details on Six areas of application?
Odyssey demonstrated the model's performance across the following domains: Robotic arms. With dozens of hours of demonstrations, the model carried out manipulation tasks and recovered from errors that weren't present in the training data — for example, a missed object grasp. Humanoid robots. Flexion, a company specializing in humanoids, built control policies on top of Odyssey-3 that handled lighting changes that caused baseline models to fail. Driving control. Odyssey tested the model's ability to drive a car on real roads in India; the policy was trained on roughly 20 hou
What are the key details on Funding and the competitive landscape?
Odyssey was founded by veterans of the self-driving car industry. In June, the company raised funding at a $1.45 billion valuation; investors include Amazon and AMD Ventures. The company operates in a rapidly growing world models sector. AMD recently agreed to acquire Fei-Fei Li's company for $8.2 billion in stock. Another company, founded by former DeepMind researchers who worked on world models, was in talks to raise $700 million. Nvidia offers its own open-source world model. As a result, Odyssey finds itself competing against formidable players: Nvidia (already embedded
Delegate Deep Research, Coding & Workflows to an AI Agent
Deploy autonomous multi-step execution with NomadNet Agent (agent.nomadnet.ai) — automated web research, data synthesis, code generation, and complex task orchestration.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.