88Z AI, 8 July — BAAI has released WuJie·RoboBrain Orca, a multimodal representation world model. The model utilizes 125,000 hours of video for unconscious learning and combines this with 160 million event annotations for conscious learning to construct a unified latent representation space of the world. It outperforms baseline models of similar scale in downstream tasks such as text generation, image prediction, and embodied action generation.
At a glance
- The model constructs a unified latent representation space through unconscious video learning and conscious event annotation learning
- It outperforms same-scale baselines in text generation, image prediction, and embodied action generation tasks
The update
BAAI has officially launched WuJie·RoboBrain Orca, a new multimodal representation world model. This model adopts a unique hybrid learning strategy, first undergoing unconscious learning using 125,000 hours of video data, and then engaging in conscious learning combined with 160 million event annotations to construct a unified latent representation space of the world. In practical application tests, WuJie·RoboBrain Orca has demonstrated strong performance advantages. In multiple downstream tasks including text generation, image prediction, and embodied action generation, the model outperforms baseline models of similar scale, validating the effectiveness and advancement of its architecture.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 智源研究院Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.