Lingchu Intelligence Details Psi-R2.5 Data Work: Reversing World Model Psi-W0 to Align Human and Robot Actions

One question has long dogged embodied AI: humans work with their hands every day, but how do those skills actually get transferred to robots? Lingchu Intelligence's latest disclosure pushes the problem one step further, arguing that putting human data into a model does not mean a human's operating experience can be used by a robot directly. Building on 100,000-hour-scale data collection, its Psi-R2.5 model continues to improve data quality and human-robot alignment methods while advancing adaptation to concrete tasks.

The difficulty is that human and robot motions do not naturally correspond. Human hands and robotic hands differ in shape, joint structure and the friction conditions when contacting objects, and hand-pose estimation carries its own error, so a motion a person performs smoothly may not succeed once converted into robot joint movement. Lingchu calls data in which humans and robots perform similar tasks that correspond mainly in task meaning weakly paired data. What it wants instead is strongly paired data: aside from the different bodies, the scenes basically match, the action sequences correspond frame by frame, and the robot-side motion can be replayed successfully on real hardware.

Such data is hard to collect directly. Having a person and a robot each perform the task separately rarely guarantees frame-by-frame correspondence in scene and timing, and simply copying human trajectories may not work either. In April this year, Lingchu tried pairing Psi-R2 with the world model Psi-W0, using trajectory rollout and reinforcement learning to turn human operations into executable robot data, but that pipeline involved coordinating multiple models and policy optimization, making data production heavy.

This time Lingchu reversed the direction: it uses Psi-W0 in reverse. Real robot images and actions are already usable data, so starting from them and generating matching human-operation data yields a new set of paired samples. Those high-quality pairs train an end-to-end converter whose input is human operation data and whose output is matched robot images and actions. Importantly, the converter is not a policy model that decides what the robot should do next; it is a video-editing model with action output.

How does the company judge whether the conversion is useful? It set up two forms of validation: replaying the converted trajectory directly on the robot to see whether the same task can be completed, and using the converted data for post-training to see whether the resulting model can complete the relevant task. The bar shifts from whether the generated video looks right to whether the data can support real operation. According to Lingchu, the converted data has been validated in both tests, though some particularly complex tasks are still not completed directly, and data conversion does not mean every task is solved.

Data quality itself was also reworked. Psi-R2.5 reduces the repeated stacking of the same task, increases task diversity, and uses an automatic annotation pipeline to break tasks into finer atomic actions before labelling and review, on the logic that data value cannot be measured in hours alone but by how many task types it covers and how much of it a robot can actually use.

Another direction is in-context learning. A new demonstration does not necessarily have to go through training again; it can be handed to an already trained model as a reference for the current task, with the robot updating no parameters and instead inferring what to do and how from contextual cues. This connects to Psi-R2.5's two-layer architecture, where an upper model decomposes long-horizon instructions into subtasks and a lower model combines subtasks with current observations to output concrete trajectories. Lingchu cautions that what it has shown is ICL capability in specific tasks, with more implementation details still to come.

The relationship between pretraining and post-training is being reframed as well. In its latest tech blog, Lingchu states that pretraining does not make post-training disappear immediately: real deployments involve highly customized object types, specifications and work rhythms, so at this stage the point of training a foundation model is to reduce how much data post-training needs, thereby lowering adaptation and deployment costs. In a phone-box assembly task, a post-training framework combining human involvement with reinforcement learning reached roughly 99% task success after several iterations in one to two working days, and data from failures at customer sites is collected for the next round of improvement.

Turning human demonstrations into data and prompts a robot can use points to a more practical path to deployment: customers do not have to train each new task from scratch, but can get a robot working with a demonstration plus a small amount of post-training data. What is worth watching next is how well this strongly paired data conversion holds up on harder tasks, and whether in-context learning can extend from specific tasks toward more open operating scenes.