Odyssey / Odyssey-3
- representation only
- 2026
- World action model
Model cardPaper ↗
Odyssey-3 is Odyssey's foundation world model: an autoregressive multi-step video diffusion transformer, with a distilled few-step variant for real-time use. The lab trains an action decoder on its representations to control machines.
Key facts
Rows below are the register record in lib/models.ts. A dash means the value is not confirmed from a primary source. Classification follows Reuss, NVIDIA Technical Blog (15 June 2026) and the model's own paper.
| Name | Odyssey-3 |
|---|---|
| Lab | Odyssey |
| Year | 2026 |
| Paradigm | representation only |
| Backbone | — |
| Source | https://odyssey.systems/meet-odyssey-3 |
What the paper claims
- Odyssey announced Odyssey-3 on 8 October 2026 as a foundation world model: a learned dynamical system implemented as an autoregressive diffusion transformer. Developers use it both to simulate environments and to train policies for physical systems. Odyssey, Meet Odyssey-3 (8 October 2026)
- The architecture is a multi-step video diffusion transformer, extended autoregressively with teacher forcing and causal masking, then distilled with distribution-matching and adversarial post-training into a few-step model for real-time interaction. Odyssey, Meet Odyssey-3 (8 October 2026)
- Training data, as stated on the lab page, mixes internet video with time-localized, schema-verified event annotations; gameplay recordings with time-aligned keyboard and mouse inputs; and simulated rigid-body interactions with captions and metadata. No parameter count or licence is given. Odyssey, Meet Odyssey-3 (8 October 2026)
- Access is a research preview named Odyssey-3 Flash, plus a contact form for physical-AI developers. The announcement does not mention open weights. Odyssey, Meet Odyssey-3 (8 October 2026)
- For control, Odyssey trains an action decoder or policy on paired observations and actions so that the world model's representations become machine-specific controls. A driving policy on 20 hours of data kept the Odyssey-3 backbone frozen and predicted waypoints from those visual representations. Odyssey, Meet Odyssey-3 (8 October 2026)
What it evaluates
- With tens of hours of arm demonstrations, Odyssey reports completed manipulation tasks and recovery behaviours that were absent from those demos, including reorienting a gripper after a missed grasp and retrieving a dropped object from an unusual position. Odyssey, Meet Odyssey-3 (8 October 2026)
- Flexion built humanoid policies on Odyssey-3. Odyssey says those policies exceeded the tested VLA baselines under environmental changes, including lighting changes that caused those baselines to fail. The post does not name the VLAs or publish a score table. Odyssey, Meet Odyssey-3 (8 October 2026)
- A driving policy trained on 20 hours of driving data, backbone frozen, is reported to have driven closed-loop on roads in India. An early multi-sensor experiment generated three front-camera views together after 100 training steps. Odyssey, Meet Odyssey-3 (8 October 2026)
- Odyssey's 15 September 2026 post describes the 20 hours of driving data as simulated, and reports that on real roads, sim-only policies travelled about 77% as far between safety-driver interventions as policies trained on real footage. Odyssey, Introducing Odyssey-3 (15 September 2026)
- Odyssey claims Odyssey-3 Pro scores 66.1 on Physics-IQ Verified video-to-video, the highest reported, and 54.7 on image-to-video. The DeepMind/Anates score table lists Odyssey-3 Pro + best-of-8 at 66.10 (top V2V row, 8 October 2026 entries) and 54.69 I2V, 0.01 behind FLUX 3 [large] + best-of-N at 54.70. A four-run Pro entry without best-of-8 is 63.37 ± 0.63 V2V. Physics-IQ Verified score table, google-deepmind/physics-IQ-benchmark
- The hosted Anates leaderboard ranks by net improvement versus track mean, not the raw Physics-IQ score. On that ranking, FLUX 3 [large] is first and Odyssey-3 Pro is second. Physics-IQ Verified leaderboard (Anates Labs)
- WorldMark is Odyssey's own evaluation. Using the benchmark's captions and the mean of 13 metric scores, Odyssey says Odyssey-3 ranked first on three of four splits: first-person stylized (published mean 79.0), third-person real, and third-person stylized (published mean 76.3). Odyssey, Meet Odyssey-3 (8 October 2026)
On this register
Representation only. Uses the video backbone's internal features and skips generating frames. That axis is defined in what is a world action model? and in Reuss, NVIDIA Technical Blog (15 June 2026).
Other representation only models on the register: Fast-WAM.
The backbone field is unconfirmed, so this card does not claim a video-model family.
Limits the authors state
- Odyssey says WorldMark measures properties of generated worlds, and that applying the model to a physical system still requires evaluating the behaviours that matter for that machine and its tasks. Odyssey, Meet Odyssey-3 (8 October 2026)
- Odyssey's Physics-IQ footnote says averaged scores use four runs and that best-of-8 uses one run with the same prompts. The official table's rules say one run is enough to appear, but a SOTA claim on i2v or v2v requires a standard deviation across four runs. Physics-IQ Verified score table, google-deepmind/physics-IQ-benchmark
- The 15 September post treats consistency across robots and environments as an open question, and says Odyssey is working with Poke & Wiggle to test transfer across bodies, viewpoints and controls. Odyssey, Introducing Odyssey-3 (15 September 2026)