Definition

Orca is a general world foundation model from baai that learns a unified world latent space via Next-State-Prediction rather than isolated next-token, next-frame, or next-action objectives. Downstream readouts expose text generation, image prediction, and embodied robot control from a frozen backbone.

Key Points

  • Architecture: Frozen qwen3.5 VLM encoder; unconscious learning (continuous video) + conscious learning (language-described events, VQA)
  • Pre-training scale: 125K hours video, 160M event annotations
  • Readout heads: Qwen3.5 language head (text); Stable Diffusion 3.5 adapters (image); Action Expert module (robotics, trained from scratch)
  • Robotics claim: Orca-4B matches Physical Intelligence π0.5 across five manipulation tasks despite zero action labels during base pre-training; only 200 real-world recordings per task for Action Expert fine-tuning
  • Error recovery: Paper examples show Orca retrying failed grasps where π0.5 repeats failures
  • Project: orca-wm.github.io; arXiv 2606.30534

Sources