Psi-R2.5

Psi-R2.5 is an embodied AI foundation model developed by the Chinese company PsiBot (灵初智能), unveiled in September 2026 as an evolution of Psi-R2. Instead of scaling data volume further, the team rebuilt the entire training pipeline around data quality — the central notion being "strong pair data", human–robot demonstration pairs aligned frame by frame.

The architecture has two layers. The upper layer, built on a QwenVL3.5-4B backbone, processes vision-language instructions and decomposes them into subtasks using metadata context (memory, soft prompts) and reinforcement learning value signals. The lower layer, built on a Wan2.2-IT2V-5B backbone, generates robot action trajectories from the upper layer's output and the current observation.

The direction of data generation is unusual: rather than translating human footage into robot data, PsiBot uses the Psi-W0 world model to inversely generate human-hand data from robot trajectories, then trains a video-editing-style transformer (not a policy) that maps human-hand input to aligned robot output. Correctness is validated by direct replay on the robot and by downstream fine-tuning success. Psi-R2.5 shows zero-shot task generalization through in-context learning — a single human demonstration, converted into robot format, is enough to adapt to a task without weight updates.