BAAI released Orca world foundation model. Models abstract world states instead of next-token/action. Qwen3.5 frozen core with swappable output heads. Action Expert trained on 200 recordings per task.
Five manipulation tasks on two-armed humanoid: matches Physical Intelligence π0.5. V-JEPA 2.1 and Qwen3.5 with same Action Expert fall behind. Orca shows error recovery advantage in paper examples.