Odyssey-3 tests a video world model as a robot backbone
Odyssey-3 is a video world model that Odyssey says can also run robot arms, humanoids and cars. We check the claimed robot results against Physics-IQ Verified.
Odyssey-3 is a foundation world model from Odyssey: an autoregressive, multi-step video diffusion transformer that predicts how a scene will change given what it has already seen and what an agent does next. Odyssey says the same weights can also be the backbone for robot action — an action decoder or small policy trained on paired observations and actions, in some cases with the video model left frozen.
On 8 October 2026 the lab posted a research preview, Odyssey-3 Flash, and a contact form for developers. The post does not name a parameter count, a licence, or open weights. That is the row now on the register.
Can a video world model be the thing a robot acts from?
That is the world action model bet, asked from the other side. Most rows on this register start as policies and borrow a video backbone. Odyssey-3 starts as a world model. The lab then trains a decoder that reads its representations and emits controls. The driving example keeps the backbone frozen, which is the same move Fast-WAM makes at inference: use the video model's features, do not wait to render a future. It is closer to DreamDojo — a world model offered as a foundation — than to DreamZero, which denoises future frames and actions together.
The robot results Odyssey published
Three claims sit in the 8 October post.
Arms. After "tens of hours" of robot-arm demonstrations, Odyssey says the system completed manipulation tasks and produced recovery behaviours that were not in those demos: reorienting a gripper after a missed grasp, and retrieving a dropped object from an unusual position.
Humanoids. Flexion trained humanoid policies on Odyssey-3. Odyssey says those policies beat the vision-language-action baselines Flexion tested under environmental changes, and kept working under lighting changes that made those baselines fail. The post does not name the VLAs or publish a score table, so we do not attach those results to any named row on this register.
Driving. A driving policy trained on 20 hours of driving data, with the Odyssey-3 backbone frozen, predicted waypoints from the model's visual representations and ran closed-loop on roads in India. The lab's 15 September post describes that 20 hours as simulated driving data, and reports a separate comparison: on real roads, policies trained only in simulation travelled about 77% as far between safety-driver interventions as policies trained on real footage.
A fourth, early experiment: after 100 training steps on the front three cameras of an autonomous-driving dataset, a checkpoint generated the three views together.
Physics-IQ Verified
Odyssey says Odyssey-3 Pro scores 66.1 on Physics-IQ Verified video-to-video, "the highest reported score", and 54.7 on image-to-video. Physics-IQ asks a model to continue videos of real physical experiments — fluids, optics, solids, magnetism, thermodynamics — and compares the continuation with what happened.
We checked the public boards.
The score table in the Physics-IQ Verified repository lists Odyssey-3 Pro + best-of-8 at 66.10 on video-to-video, the top V2V row among the 8 October 2026 entries. The same best-of-8 row for image-to-video is 54.69, 0.01 behind FLUX 3 [large] + best-of-N at 54.70. Odyssey's post rounds the I2V figure to 54.7. A four-run Odyssey-3 Pro entry without best-of-8 sits at 63.37 ± 0.63 on V2V.
Odyssey's own footnote says averaged scores use four runs, and that best-of-8 uses one run with the same prompts. The repository's rules say one run is enough to appear on the board, but claiming state of the art on i2v or v2v requires a standard deviation across four runs.
The hosted Anates Labs leaderboard ranks models by net improvement versus the track mean, not by that raw score. On that ranking, FLUX 3 [large] is first and Odyssey-3 Pro is second.
WorldMark is Odyssey's own evaluation
Odyssey also reports WorldMark, which it describes as measuring control-following, visual quality and world memory. It says that in its evaluation, using the benchmark's own captions and the mean of 13 metric scores, Odyssey-3 ranked first on three of four splits: first-person stylized, third-person real, and third-person stylized. The published means we can read off that post are 79.0 on first-person stylized and 76.3 on third-person stylized. This is Odyssey scoring Odyssey. We have not found a public WorldMark leaderboard independent of that post.
What Odyssey itself limits
The 8 October post says WorldMark scores measure properties of generated worlds, and that using the model on a physical system still needs evaluation of the behaviours that matter for that machine and its tasks. 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.
Access is a research preview. Developers are asked to get in touch. No open weights, parameter count or licence appear on either lab page.
On this register, Odyssey-3 is filed as representation only. That is a classification of how actions come out — from the backbone's features — not a claim that Odyssey-3 is already a deployed robot policy. Whether a video world model can carry that load is the argument the lab is making, and the reason the row is here.