Psi-R2

A world-action model from the Chinese startup PsiBot, pretrained on roughly 100,000 hours of human manipulation data.

Psi-R2 is an embodied model released on April 10, 2026, by the Chinese startup PsiBot (灵初智能). The company describes it as the first 'world-action model' pretrained at the scale of 100,000-plus hours of human manipulation data: a model that takes an image and a language instruction and jointly predicts future video frames and the robot's actions. The data comes from human demonstrations captured with the company's own exoskeleton gloves, covering scenes such as industrial assembly, everyday chores, and object grasping; PsiBot also open-sourced an initial batch of 1,000 hours of human hand-operation data, including vision, language, joint angles, and touch. The company states that fine-tuning on fewer than 100 real-robot trajectories is enough for the model to perform long-horizon, fine-grained tasks such as assembling a phone, industrial packaging, or folding paper boxes. Alongside Psi-R2, PsiBot released an action-conditioned world model called Psi-W0, used to convert human data into robot-usable data and to run reinforcement learning on the policy inside the model. A follow-up, Psi-R2.5, released on September 28, 2026, switched to a two-level structure with a high-level VLM breaking tasks down and a lower-level video model producing the actions.

ExampleIn an official demo, after fine-tuning on fewer than 100 real-robot trajectories, the robot completes a multi-step, fine-grained task such as assembling a phone.

Also called

Psi R2, PsiR2, PsiBot Psi-R2

Related

World Action Model · PsiBot · Human Video Data · Data Glove · World Model · DreamZero

Sources

灵初智能：新一代具身模型发布，全球最大人类手部操作数据集开源 (Chinese)
灵初智能：Psi-R2.5 正式发布 (Chinese)

As of

2026-09