Odyssey reports robot-arm tasks from tens of hours of demonstrations, including recoveries that were not in those demonstrations, such as reorienting a gripper after a miss. Flexion, they say, trained humanoid policies on Odyssey-3 that beat the VLA baselines they tested when the environment and the lighting changed. A driving policy trained on 20 hours of road data in India, with the backbone frozen, predicts waypoints and drives in closed loop. An early checkpoint also generated three camera views together after 100 training steps.
The model is trained as a multi-step video diffusion transformer, extended autoregressively with teacher forcing and causal masking so it continues from what it has already generated, conditioned on actions. A distillation pass produces a few-step variant for real-time interaction. The preview lets a person prompt a world, move, and introduce an event while it is generating. Training data, as described, mixes internet video with event labels, gameplay with keyboard and mouse, and simulated rigid-body interactions.
WorldMark, in Odyssey’s evaluation, averages 13 metrics of control-following, visual quality, and memory. The lab says the model ranks first in three of four categories in its own run: first-person stylized, third-person real (79.0), and third-person stylized (76.3). Physics-IQ continues videos of real experiments in fluids, optics, solid mechanics, magnetism, and thermodynamics. Both leaderboards are Odyssey’s report.