Lingchu Intelligence Releases Psi-R2.5: Reverse-Generating Paired Data to Improve Robot Learning Efficiency

What Lingchu Intelligence is asking this time is a seemingly roundabout question: human data entering the model does not mean human operational experience can be directly used by robots—what is still missing in between? According to QbitAI, building on 100,000-hour-scale data accumulation, Psi-R2.5 continues to improve data quality and human-robot data alignment methods, and further advances specific task adaptation.

The problem lies in the differences between human and robot embodiments. The report points out that human hands and robotic hands not only look different, but also have different joint structures and friction conditions when contacting objects; combined with errors in human hand pose estimation, an action a person can smoothly complete may not necessarily succeed after being converted into robot joint motion. Therefore, letting a robot understand 'this is picking up a cola' and providing a piece of action data for 'you can pick up the cola like this' are two different things.

Lingchu calls data where humans and robots perform similar tasks and mainly correspond in task meaning 'weakly paired data'; what it wants more is 'strongly paired data': aside from different embodiments, the scenes on both sides are basically consistent, the action timing corresponds frame by frame, and the robot-side actions can be directly replayed successfully on real machines. Such data is very difficult to collect directly—having a human and a robot each perform it once makes it hard to guarantee frame-by-frame correspondence in scenes and action timing, and directly copying human trajectories may not necessarily succeed.

In April this year, Lingchu already tried to have Psi-R2 work with the world model Psi-W0, converting human operations into executable robot data through trajectory inference and reinforcement learning. But the report points out that this pipeline involves coordination of multiple models and policy optimization, making the data production process relatively heavy.

Lingchu then changed direction: the images and actions of real robot operations are already usable—if they are used as a starting point to generate matching human data, wouldn't that yield a new set of comparison samples? Following this line of thought, Lingchu reverse-uses Psi-W0, generating corresponding human-hand operation data from real robot data, and then uses these high-quality paired data to train an end-to-end converter; the converter's input is human operation data, and its output is matching robot images and actions.

The report specifically distinguishes the converter's positioning: it is not a policy model responsible for deciding what the robot should do next, but a video editing model with action output; what it does is not just replace the human hand in the frame with a robotic hand, but simultaneously obtain action data the robot can use. The report also mentions that simply recording a human-hand operation video with a phone in a daily scene can be converted through the model into corresponding robot data.

Lingchu set up two verification methods: one is to directly replay the converted trajectories on the robot to see whether it can complete the same task; the other is to use the converted data for post-training, then see whether the trained model can complete the related task. The criterion is no longer just 'does the generated video look similar,' but whether the data can support actual operation.

According to Lingchu, the converted data has been verified in the above two types of tests. But the report also gives boundaries: some particularly complex tasks are still not directly completed, and data conversion does not mean all tasks have been solved.

Following this line of thought, Psi-R2.5 also reorganized the quality of the data itself: reducing repeated accumulation of the same task, increasing task diversity, and using an automatic annotation pipeline to break tasks down into finer atomic actions before annotation and review. The report's formulation is that the value of data cannot be measured only by hours; it also depends on how many kinds of tasks are covered and how much the robot can use.

If a segment of human operation can already be converted into a reference that robots can use more easily, it may also have another use: not only entering the training set, but also becoming a prompt for the current task. This means there is no need to write all requirements into a long instruction; instead, give the robot an example—'follow this.'

This is the in-context learning (ICL) that Lingchu further explores in Psi-R2.5. In the tasks it demonstrates, the robot does not need to update model parameters, but instead infers what to do and how to do it based on clues provided by the current context. The key change is: a new demonstration does not necessarily have to go through the training process again to affect robot behavior; it can also serve as a reference for the current task, given to an already trained model to use.

The difficulty still lies in the fact that the demonstration comes from a human, while the executor is a robot. Lingchu's idea is to use the earlier strongly paired data model to convert human demonstrations into corresponding robot data, then input it into the model as a prompt. The same data conversion capability thus has two uses: preparing training materials, and helping the robot understand the demonstration in front of it. This also connects with Psi-R2.5's two-layer architecture: the upper-layer model is responsible for decomposing long-horizon tasks, splitting longer instructions into subtasks; the lower-layer model combines subtasks and current observations to output specific operation trajectories. The report also cautions that 'no need to retrain' does not mean the model has not been trained before, nor does it mean that casually shooting a video can make the robot complete any work; what is currently demonstrated is ICL capability in specific tasks, and more implementation details remain to be announced.

Pre-training has already used 100,000-hour-scale data, so why does the robot still need to keep training at the customer site? In Lingchu's view, real deployment often faces strongly customized requirements for item types, specifications, and work tempo; in the short term, one cannot expect the foundation model to enter all customer sites directly without any adaptation, and post-training remains a necessary step.

In its latest Tech blog, Lingchu clearly points out that pre-training does not make post-training disappear immediately. From a deployment perspective, the significance of training a foundation model at this stage should be to reduce the data needed for post-training, thereby lowering the cost of adaptation and deployment.

Phone box assembly is a concrete example: this task involves multiple steps and also has relatively high requirements for operational precision. According to Lingchu, through a post-training framework combining human participation and reinforcement learning, after several rounds of iteration, the task success rate reached about 99%, taking 1-2 working days. Training does not end at the customer site either: wherever the robot fails, the relevant data is collected back for the next round of improvement; the team says this solution has been used in actual customer sites to handle complex situations encountered in operations. The report says Psi-R2.5 is applying the capabilities accumulated in pre-training to specific work, with the goal of letting customers spend less time and less cost to get robots up and running.