Meet Odyssey-3: Our Most Powerful Foundation World Model

TL;DR

Odyssey’s follow-up to the September Introducing Odyssey-3: A General-Purpose Physical Intelligence launch pins Odyssey-3 to concrete benchmark numbers: Odyssey-3 Pro reports 66.1 on the Physics-IQ Verified video-to-video track as the leaderboard-top score, 54.7 on image-to-video, and 1st-place on 3 of 4 WorldMark splits (first-person stylized, third-person real, third-person stylized) under WorldMark’s own captions on the mean of 13 metrics. The post also unlocks a research-preview environment-generation surface (first- and third-person navigation + independent camera, text-prompted environments) and documents the training recipe at a higher level than before — multi-step video DiT taught with temporally-resolved prompts and controls, extended autoregressively via teacher forcing with causal masking, then distilled with a distribution-matching + adversarial pipeline into a few-step real-time variant. Robot-arm, humanoid, driving, and multi-sensor downstream deployments are reiterated with new specifics (three-camera joint generation after only 100 training steps for autonomous-driving sensor arrays) under the same frozen-backbone + small action decoder recipe already established in the September post.

Key claims

Odyssey-3 Pro sets a new state of the art on Physics-IQ Verified’s video-to-video benchmark at 66.1, the highest reported score on the leaderboard at the time of publication [§Odyssey-3 is State of the Art on Physics-IQ].
Odyssey-3 Pro additionally scores 54.7 on Physics-IQ image-to-video; Odyssey-3 (the smaller variant) is positioned as improving the physical-accuracy-vs-generation-cost tradeoff rather than topping the absolute scoreboard [§Odyssey-3 is State of the Art on Physics-IQ].
Cost numbers on the Physics-IQ leaderboard assume $1 per MI355X GPU-hour (excluding prompt-rewriting fees) with Odyssey-3 at 832×480 and Pro at 1280×720; scores average 4 runs while best-of-8 uses 1 run with the same prompts [§Odyssey-3 is State of the Art on Physics-IQ caption].
On WorldMark — a 13-metric suite covering control-following, visual quality, and world memory — Odyssey-3 ranks first on first-person stylized (77.2 vs #2 LingBot-World 77.0), third-person real (79.0 vs #2 HY-World 1.5 76.9), and third-person stylized (76.3 vs #2 HY-World 1.5 75.1), using the benchmark’s own captions and the mean of its 13 reported metric scores [§WorldMark].
Odyssey-3 does not lead first-person real on WorldMark: Lyra 2.0 ranks 1st at 84.4, AlayaWorld 2nd at 83.0, Odyssey-3 3rd at 80.6 [§First-Person Real].
The research preview exposes an interactive environment-generation surface: text-prompted environment creation with real-time response to user actions / events / observations, with first-person navigation, third-person navigation, and independent camera-movement modes [§Odyssey-3 Generates Environments in Real Time].
Training-pipeline disclosure (first filed for Odyssey): base model is a multi-step video diffusion transformer using temporally resolved prompts and controls, extended autoregressively via teacher forcing with causal masking (so each future state is conditioned on preceding observations + action inputs), then post-trained with a distribution-matching + adversarial distillation pipeline into a few-step real-time variant [§How We Built Odyssey-3].
Training-data composition disclosure (also first filed): internet video with time-localized schema-verified event annotations, gameplay recordings with time-aligned keyboard+mouse inputs, and simulated rigid-body interactions with captions + metadata [§How We Built Odyssey-3].
Robot-arm results repeat the September blog’s qualitative recovery-behavior claims under the same frozen-WFM + action-decoder recipe (reorienting a gripper after a missed grasp, retrieving a dropped object in an unusual position), without new quantitative numbers [§Odyssey-3 Can Control Robot Arms].
Humanoid results via the Flexion partnership are reiterated: policies trained on Odyssey-3 “exceeded the performance of the tested VLA baselines under environmental changes and continued to perform tasks under lighting changes that caused those baselines to fail” [§Odyssey-3 Can Power Humanoids].
Driving deployment reiterates the frozen-backbone + waypoint-policy recipe at 20 h of real-world Indian-roads data; the earlier blog’s “77% of real-trained between-intervention distance” quantitative claim is dropped in favor of qualitative close-loop driving footage [§Odyssey-3 Can Drive Vehicles].
New deployment claim: adapted to generate observations for particular sensor arrangements — the front-three-camera arrangement of an autonomous-driving dataset produces plausibly synchronized three-camera driving sequences after only 100 training steps from an Odyssey-3 checkpoint [§Odyssey-3 Can Generate Multi-Sensor Data].

Method

Odyssey-3 is positioned as a learned dynamical system implemented as an autoregressive diffusion transformer. The training stack is now disclosed at method-level granularity for the first time: a multi-step video diffusion transformer is trained with temporally resolved prompts and controls that guide how a scene unfolds under action inputs; this base is extended autoregressively via teacher forcing and causal masking so that the model learns to continue from preceding observations and predict future states conditioned on action inputs; a post-training pipeline then applies distribution-matching + adversarial distillation to produce a few-step distilled variant that supports real-time interaction. Training data combines three sources — internet video with time-localized schema-verified event annotations, gameplay with time-aligned keyboard/mouse inputs, and simulated rigid-body interactions with captions + metadata — explicitly to connect observations with descriptions of what happens and (where available) the actions that produced it.

Downstream adaptation follows the recipe established in the September post (Introducing Odyssey-3: A General-Purpose Physical Intelligence) and reiterated here: freeze the pretrained backbone and attach a small action decoder or policy trained on paired observations + actions. The driving section explicitly re-states “keeping the Odyssey-3 backbone frozen” and training “only” a waypoint policy on 20 h of real-world Indian-roads data; the robot-arm and humanoid sections use the same decoder framing without re-stating “frozen” explicitly. Multi-sensor extension (front-three-camera autonomous-driving output) is described as adapting an Odyssey-3 training checkpoint rather than from-scratch retraining, with the three-camera co-generation emerging after 100 training steps.

Results

Headline benchmark numbers (reported as of 2026-10-07 on the Physics-IQ Verified leaderboard):

Physics-IQ Verified [§Odyssey-3 is State of the Art on Physics-IQ]:

Odyssey-3 Pro video-to-video: 66.1 (reported as the highest leaderboard score).
Odyssey-3 Pro image-to-video: 54.7.
Odyssey-3 (non-Pro) positioned on an improved physical-accuracy-vs-cost tradeoff frontier; absolute scores not called out in the text.

WorldMark (mean of 13 reported metrics per split, using WorldMark’s own captions) [§WorldMark]:

First-Person Stylized (11 models): Odyssey-3 77.2 (#1), LingBot-World 77.0, AlayaWorld 76.7, Lyra 2.0 75.9, Matrix-Game 3.0 75.6, HY-World 1.5 75.2, Matrix-Game 2.0 73.5, SANA-WM 73.0, DreamX-World 70.6, Yume 1.5 63.3, HY-GameCraft 1.0 55.6.
First-Person Real (11 models): Lyra 2.0 84.4 (#1), AlayaWorld 83.0, Odyssey-3 80.6 (#3), Matrix-Game 3.0 80.1, LingBot-World 79.6, Matrix-Game 2.0 77.1, HY-World 1.5 77.0, SANA-WM 75.2, DreamX-World 72.2, Yume 1.5 67.6, HY-GameCraft 1.0 59.5.
Third-Person Real (6 models): Odyssey-3 79.0 (#1), HY-World 1.5 76.9, SANA-WM 74.8, DreamX-World 71.5, LingBot-World 69.5, Matrix-Game 2.0 68.6.
Third-Person Stylized (6 models): Odyssey-3 76.3 (#1), HY-World 1.5 75.1, SANA-WM 72.8, DreamX-World 70.1, Matrix-Game 2.0 65.5, LingBot-World 60.8.

Robotics / embodied (reiterated from the September blog, no new numeric results): emergent recovery behaviors on robot arms (reorienting gripper after missed grasp, retrieving dropped object), humanoid lighting-generalization wins vs tested VLA baselines via the Flexion partnership. Driving reiterated with 20 h of real-Indian-road data under a frozen backbone — the September blog’s “~77% as-far between-intervention” quantitative headline is not repeated in this post.

New qualitative result: three-camera synchronized autonomous-driving video sequences from an Odyssey-3 training checkpoint after only 100 training steps on a target sensor arrangement [§Odyssey-3 Can Generate Multi-Sensor Data]. No quantitative sensor-fidelity metrics (no camera-consistency score, no comparison to a baseline multi-view video generator) are reported.

No head-to-head against Cosmos 3, Genie 3, Runway GWM-1, or Veo 3 is disclosed on either benchmark. The Physics-IQ ranking is relative to the Physics-IQ Verified leaderboard; the WorldMark ranking is specifically “in our evaluations” using WorldMark’s own captions. Research preview available as of publication.

Why it’s interesting

This is Odyssey’s benchmark-landing follow-up to the general-purpose-physical-intelligence framing of Introducing Odyssey-3: A General-Purpose Physical Intelligence, and it is the first filed WFM-cluster entry where a closed-flagship interactive-rollout WFM publishes a leaderboard-first on Physics-IQ (previously the strongest filed Physics-IQ datapoint was Towards World Simulator: Crafting Physical Commonsense-Based Benchmark for Video Generation (PhyGenBench)‘s “best of 8 models scores only 24.1” baseline, with VJEPA-WMReward boosting Inference-time Physics Alignment of Video Generative Models with Latent World Models MAGI-1 to 62.0 as a post-hoc reward-gradient reranking rather than a native-model score). The 66.1 V2V number — if independently reproducible — pushes the “pretrained video generators are not law-grounded world simulators” ceiling that Video Generation Benchmarks has tracked across PhyGenBench, VideoPhy-2, Physics-IQ, and Apple-π. The disclosed training recipe (multi-step DiT → teacher-forced AR extension → DMD + adversarial distillation into a few-step real-time variant) is the recipe-template the Diffusion Distillation cluster now sees repeatedly in interactive-WFM releases (Self-Forcing++, DMD2, LongLive-2.0, Causal Forcing), making Odyssey-3 a public-facing instance rather than a new design pattern. The split result on WorldMark — Odyssey-3 leads 3 of 4 splits but loses first-person real to Lyra 2.0 at 84.4 vs 80.6 — is a useful counter-signal to any “Odyssey tops every benchmark” reading, and the fact that LingBot-World (LingBot-World 2.0 / LingBot-World-Infinity — Infinite Worlds with Versatile Interactions) and AlayaWorld (AlayaWorld: Interactive Long-Horizon World Modeling) sit in the top 3 across every first-person split without this blog mentioning them is itself evidence that the open-weights interactive-WFM pole is now competitive on the same benchmark as the closed-flagship pole. Research-preview access (vs. the September blog’s “coming weeks”) lets external evaluators independently reproduce the Physics-IQ claim; weights are still not open.