Awomo Takes the Top Two Spots on the RoboCasa365 Leaderboard
Awomo’s two world models have secured the TOP TWO positions among 18 evaluated models on the RoboCasa365 multi-task learning leaderboard. Awomo-0.5I ranks first with an overall score of 64.4, followed by Awomo-0.5V in second place with 62.5.
Evaluating Robotic Manipulation Across 50 Tasks
RoboCasa365 evaluates robotic manipulation in simulated kitchens, from individual skills to multi-step activities. Its multi-task benchmark covers 50 tasks, testing basic manipulation, task sequencing, and generalization to tasks unseen during pretraining.
| Model | Overall Score | Rank |
|---|---|---|
| Awomo-0.5I | 64.4 | 1st |
| Awomo-0.5V | 62.5 | 2nd |

Two Approaches to World Modeling
Awomo-0.5I uses image generation to predict key future visual states, connecting current observations and task goals with anticipated outcomes to guide action decisions.
Awomo-0.5V uses video generation to capture how objects and environments change over time, deriving executable robot actions from predicted visual sequences.
The RoboCasa365 results demonstrate both models’ ability to handle a range of kitchen tasks, from individual actions to multi-step sequences. Awomo-0.5I achieved 45.2% on Composite-Unseen, while Awomo-0.5V reached 60.4% on Composite-Seen, highlighting their performance on tasks unseen during pretraining and learned multi-step activities, respectively.
These results mark progress toward robots that can combine individual skills to complete more complex kitchen tasks. Awomo will continue to improve multi-step execution and generalization to unfamiliar tasks, while validating these capabilities beyond simulation.
Explore the RoboCasa365 leaderboard, and find Awomo-0.5 on Hugging Face and GitHub.
Scores and rankings are based on the RoboCasa365 leaderboard updated on October 13, 2026.