Two years ago, a video of a robot shooting went viral on the Internet.
In the video, the mechanical arm with one hand on the end of the stick, one hand on the handle of the stick, rolls the stick over and throws it into the bowl, which looks like a skilled kitchen assistant.
However, it is not an easy task for a robot to learn. Researchers must first personally manipulate a robot to complete a task, repeatedly collect real-world samples, and then gradually learn to perform autonomously through simulation learning. For a fixed task, about 50 such samples are collected.
Nowadays, early intelligence has demonstrated another possibility for robot learning in similar tasks. In this group of presentations, humans only need to demonstrate once, without re-training or fine-tuning the model, and robots can use this operation as a context similar to Prompt to try to complete new tasks.
Following a three-step process of human demonstration "prepare dishes, snacks, frogs, toads, and lobsters", Psi-R2.5 reproduces the model in sequential order through In-Context Learning without having to re-train the model.
Similarly, two years ago, the focus was on robots "would" and now Spirit is more focused on robots "how to learn".
It's almost subconsciously done for people to pick up and roll up supplies. Could it be possible to jump over and over again to collect real machine data and directly use human experience to train robots?
The Spirit was already moving along this path.
Released in April, Psi-R2 and Psi-W0 used 100,000 hours of human data to pre-train and further optimize human trajectories through Psi-W0 and reinforcement learning into trajectories that robots can execute.
Now, Intel has officially released a new generation body-based model, Psi-R2.5.
This time, instead of focusing on "100,000 hours of human data", the team started working on two new issues after the data was scaled up: which data is really valuable and how to get that human data into the robot training process more smoothly.
Project presentation and full technology blog: Scaling Pair Data for Embodied Intelligence
There's a crucial change in data conversion, and since it's difficult to convert human data to robot data, can we start with real robot data that could actually be executed?
The answer lies in Pair Data.
Scaling Pair Data：
Let human data really get into robot training
Why can't we use human video directly to train robots?
On the one hand, human and mechanical hands have differences in vision, camera parameters, environmental changes, and so on; on the other hand, between human action and real robot data, there are also hand gesture estimates, joint structural mobility errors, and physical parameter errors such as frictional force, or Dynamic Embodiment Gap.
As previously mentioned, the previous generation of Psi-W0 has been able to optimize human trajectories into executable robot trajectories, but each data still needs to be refined through World Model rollout and reinforcement learning.
Psi-R2.5 has changed the way Pair Data is built.
Initially, the data collected only according to the same mission semantics is called "weak Pair Data"; if the image is essentially consistent outside the body, and the action can be replayed directly on the robot, it is called "strong Pair Data".
Powerful Pair Data
Weakness of Pair Data
The team started from real robot data and reversed to generate corresponding human-handed data.
It's an interesting step: finally, it's clear that if you want to turn human data into robot data, why do you have to re-create human data?
The key is that the real robot data itself already contains images and actions that can be used directly. The team starts with these robot data and generates human-to-human matching data before getting a batch of human-to-human matching robot data.
With these powerful Pair Data sets, the team has further trained end-to-end human-to-robot data conversion models with action, so that new human-operated data can be directly synchronized with robot data.
One way to verify the conversion is to replay the converted robot data directly to the real machine, and the other is to use the data to continue training the model.
In the end, people only need to shoot an operating video with a normal phone or camera to convert it into data from different robots through this pipeline.
Compared to the previous generation, this really shortened the gap between "getting a piece of human data" and "getting data that robots can actually use".
From human demonstrations to real missions,
Just one or two working days.
But turning human data into data that robots can use only solves the problem of "how to learn".
In the face of specific scenarios, robots also need to handle a series of changes, such as task steps and the on-site environment.
As a result, an adjustment was made to the model structure of the Psi-R2.5.
Compared to the two previous generation Psi-R2 and Psi-W0 models, which had different functions, the R2.5 had a two-tiered architecture: the upper layer was responsible for dismantling long-range tasks, dividing a long instruction into corresponding subtasks; the lower layer combined current observations to produce robot-specific operational trajectories. The upper model was based on QwenVL3.5-4B, while the lower layer used Wan2.2-IT2V-5B, which were both trained using the same set of pre-training data.
However, the basic model is still difficult to achieve without additional training to respond directly to all tasks. The SKU, steps and work rhythms in different scenarios are often highly customized, so post-training is still necessary for a considerable period of time.
To this end, Lin initially developed a set of hands-on HIL ((Human-in-the-Loop) + RL post-training frameworks: a small amount of data is required to continue to fine-tune the base model; instances of failures in the deployment process will also flow back in real time for subsequent iteration.
This is an intuitive example of how mobile phones are fitted.
This is a step-by-step process that requires a high level of precision operation. If you start directly from zero with simulation learning training, the success rate is low; whereas on the basis of HIL and post-training RL, the success rate can be increased to 99% after a few rounds of iteration, and the entire process takes only 1 ton2 working days.
Scaling Pair Data solves how to turn human models into training data that can actually be used by robots, while HIL and RL allow the model to perform tasks more steadily after entering a specific site.
For temporary new tasks, Psi-R2.5 also explores another lighter approach: In-Context Learning. First, humans demonstrate and then use this trajectory as an upstream context input similar to Prompt, allowing robots to complete tasks that will not be completed until the demonstration is complete.
There is also a slight difference between the context and the language model of the robot, which is not just a text but a human instructional video.
Using the powerful Pair Data conversion capability of the previous one, human samples can be converted into corresponding robot data and then as a Prompt input model. ICL can allow robots to generate zero samples for new tasks based on text cues or physical cues without updating model parameters.
From a practical point of view, these two methods can meet different task needs: tasks that require long-term stable execution can be continued through post-training and on-the-spot data mutations; and new tasks can be more easily adapted through demonstration and literacy lessons.
After 100,000 hours,
Data scaling starts with "quantity of information".
Suppose there are 10,000 hours of data covering 100 tasks, each of which is repeated for 100 hours; another batch of data covering 100 hours, also covering 100 tasks, each of which is repeated for 1 hour.
The time difference between the two is 100 times greater, but the actual amount of mission information contained may not be that much different.
Thus, after the human data reached the 100,000-hour scale, Psi-R2.5 began to actively compress the repetitive data: reducing the amount of data on a single mission to cover as many tasks as possible on the same data scale; while using Autolabeling data pipelines to fine-tune and audit the split atomic missions.
The focus of scaling also began to shift from simply increasing the length of time to increasing the amount of information in the same data scale.
This change also extends to evaluation.
Psi-R2.5 introduced simulation evaluation during training and set up a multitasking machine evaluation set containing 50 complex tasks. It also maintains the initial state of randomized tasks during testing to observe the performance of models in combined generalization.
Psi-R2.5 multi-tasking machine evaluation, covering 50 complex tasks, and randomizing the initial state of the task.
So how do we judge that a piece of human data isn't good enough?
In fact, the standards are simple, really high-quality human data, and should be able to directly train models to complete corresponding tasks.
Two previous tests used to verify data conversion effects have also become the standard for measuring data quality: data that has been converted can be replayed on a robot, but can also be used for real post-training, and the model can be generalized to relevant tasks.
This also explains why Pair Data will be an important part of this round of upgrades to Psi-R2.5.
If a piece of human data can only tell the model "what humans do" at a semantic level, it still has limited help for robot operations; and when it is converted, it can directly replay, can bring in a training model, and it really provides the robot's most needed operating information.
From Psi-R2 to Psi-R2.5, a marked change can be seen in the data path from the beginning of the Spirit: first, to scale human data to 100,000 hours, and then to start to recalculate how valuable that information is in 100,000 hours.
Written at the end
With physical intelligence moving to scale, data costs have always been an inevitable threshold.
Every time a robot enters a new scene, learns a new task, and if it has to retrieve large amounts of real machine data and complete a round of training and adaptation, the speed of capacity expansion will soon be delayed by data production.
Human data offers another possibility. It is rich enough and more easily available, but the scale advantage only makes sense if it is actually transformed into operational data that robots can execute, train, and generalize.
From Psi-R2 to Psi-R2.5, the company has been moving in this direction. Strong Pair Data improves the quality of coordination between human data and robot data, HIL+RL compresses the cost of adapting specific tasks, and ICL further attempts to make robots less retrained when faced with new tasks.
It is worth looking at how much human data will reduce the cost of training and deploying robots in the future.