Odyssey-3 is a foundation world model designed to control diverse physical systems including humanoids, robot arms, drones, and autonomous vehicles using a shared understanding of physics. The "Physics Agent" learns world behavior from visual observation and requires minimal task-specific data, including 20 hours of simulated driving data for autonomous vehicle control. Policies for humanoid and robot arm systems used tens of hours of demonstrations to achieve operational control.

The system utilizes a pretrained world model paired with an action decoder to translate general representations into the actions of a specific machine. This design targets the task-specific data bottleneck by spending a high pretraining budget on a general model to avoid relearning knowledge for every new body. Odyssey plans a public release of the model in the coming weeks following the current preview.

Sign in to suggest edits

Key sources

  1. SOURCE@testingcatalog“Odyssey-3, a world model that it calls a Physics Agent!”x.com
  2. SUPPORT@wallstengine“learned autonomous driving from just 20 hours of simulated driving data, while robot and humanoid policies were trained with only tens of hours of demonstrations”x.com
  3. SUPPORT@testingcatalog“public release planned in the coming weeks”x.com
  4. SUPPORT@omarsar0“reusing world knowledge across machines, with a few hours of experience teaching each one how to use its controls”x.com
  5. SUPPORT@rohanpaul_ai“action decoder trained on tens of hours of robot demonstrations”x.com
  6. SOURCEmarketbrief.now
Markdown