Lingchu Intelligence Releases Embodied Foundation Model Psi-R2.5, Reworking Training Pipeline with Strong Pair Data

Two years ago, a video of a robot cooking shrimp went viral. Behind it, researchers repeatedly collected real-robot demonstrations and then used imitation learning to let the robot execute independently; a single fixed task required about 50 demonstrations. Now Lingchu Intelligence demonstrates another possibility in a similar task: a human demonstrates once, and the robot uses that operation as context similar to a prompt to attempt a new task. This contrast is the starting point for understanding Psi-R2.5—the focus shifts from whether the robot "can" to "how it learns."

Lingchu was already moving along this path. Psi-R2 and Psi-W0, released in April, used 100,000 hours of human data for pretraining, and through Psi-W0 and reinforcement learning optimized human hand trajectories into trajectories the robot could execute. But this pipeline means the Policy and World Model need repeated rollouts, and the trajectories still need further fine-tuning with reinforcement learning; the conversion process remains cumbersome.

In the newly released Psi-R2.5, the team did not continue to dwell on "100,000 hours of human data," but instead addresses two new problems that arise once data scale grows: which data is truly valuable, and how to make this human data enter the robot training process more smoothly.

Why can't human videos be directly used to train robots? The core is the Embodiment Gap. On one hand, human hands and robotic hands differ visually, and camera parameters, environmental changes, and so on create a Visual Embodiment Gap; on the other hand, between human actions and real robot data there are also hand pose estimation errors, kinematic errors from joint structure, and physical parameter errors such as friction, namely the Dynamic Embodiment Gap.

Psi-R2.5 changes the way Pair Data is constructed. Lingchu calls data collected only by the same task semantics "weak Pair Data"; if the images are essentially identical except for the embodiment, correspond frame by frame in timing, and the actions can be directly replayed on the robot, it is called "strong Pair Data." The team starts from real robot data and reverse-generates corresponding human hand data—the real robot data itself already contains directly usable images and actions, and after generating one-to-one corresponding human hand data, a batch of paired human-robot data is obtained first.

With this batch of strong Pair Data, the team further trains an end-to-end human-data-to-robot-data conversion model with actions, so that new human operation data can be directly aligned with robot data. There are two validation methods: directly replaying the converted robot data on the real robot, or using this data to continue training the model. The vendor claims that a person only needs to capture an operation video with an ordinary phone or camera, and this pipeline can convert it into data for different robots.