For robots to perform complex tasks as skillfully as humans, imitation learning alone is far from enough.

Today, Physical Intelligence (hereinafter referred to as PI) released its latest foundation model, π*0.6. Through an innovative Recap training method that integrates demonstration, guidance, and autonomous practice, the Visual-Language-Action (VLA) model has broken through the bottlenecks of imitation learning.

As a startup in the field of Embodied AI, PI is headquartered in San Francisco and completed over $400 million in funding last year. Technically, they remain strongly committed to the VLA approach and prioritize enabling models to execute tasks in new environments.

Co-founder Sergey Levine proudly stated that robots equipped with the π*0.6 model are already able to make coffee for everyone in the company, and their success rate in many tasks has reached nine tenths.

Making Robots More Than Partially Competent

Over the past year, many advancements in robot learning have relied on imitation learning, which trains models using human demonstration data. However, this approach only allows robots to "know half" of what needs to be done.

In real-world dynamic environments, robots may encounter a cascade of errors—starting small and growing larger—that cannot be captured in demonstration data, leading to the problem of "error accumulation."

Relying solely on imitation of demonstrations does not guarantee high success rates. Achieving 100% reliability and human-level processing efficiency is still far off.

Therefore, Pi's newly released π*0.6 model aims to address issues of success rate and efficiency in robots' actual operational tasks.

At its core, the π*0.6 model employs the Recap training method, which enables robots to go through three stages: "demonstration learning, error-correction guidance, and autonomous practice."

The model can extract effective training signals from so-called "bad data," rather than merely having robots repeat actions based on demonstrations, which would only cause them to make the same mistakes repeatedly.

During the error-correction guidance phase, remote operators take over when the robot makes an error, demonstrating how to recover from it. This targeted correction data directly addresses specific problems encountered by robots in real-world scenarios, breaking the cycle of error propagation.

(Remote operators take over to provide real-time corrections)

However, relying solely on manual error correction makes it difficult to cover all nuanced scenarios; robots still need to continuously improve through autonomous practice.

Therefore, Recap introduces a value function to solve the 'credit assignment' problem in reinforcement learning, which determines which actions led to ultimate success or failure. By predicting task completion probabilities across different scenarios, the model can identify key effective actions and error nodes, thereby reinforcing advantageous behaviors and avoiding erroneous operations.

Value function visualizations show that when a robot successfully completes a task, the curve rises sharply, but drops significantly when errors occur. These fluctuations help the model extract signals from its own experiences.

From imitation to error-correction guidance and then to autonomous practice, robots avoid situations where small mistakes accumulate into large ones during learning. In this phase, 'bad data' becomes a major driving force, with reinforcement learning at its core.

How does it perform in practical applications?

As another masterpiece of PI in VLA models, π*0.6 is built on a five billion parameter vision-language model, paired with an action expert module, supporting multi-dimensional conditional inputs such as text instructions, execution quality, and action advantage.

In practical applications, PI focused on three scenarios: making coffee, folding clothes, and assembling cardboard boxes. Each task presents significant challenges, leading to low success rates when performed repeatedly.

For instance, making coffee is a complex, long-threaded task that involves pouring milk, determining whether the machine has started or finished operating, and offering options for post-task robot cleaning—a process with an extremely wide scope.

When folding clothes, robots face the diversity of materials, requiring different postures and force application methods to handle various garments.

Finally, in the task of assembling cardboard boxes, robots must cope with complex and precise challenges such as flat boxes sticking together, deviations in edge folding, and tape pasting.

These tasks represent current difficulties for robotic models, which struggle to balance both quality and quantity. However, after Recap training, robots powered by π*0.6 can produce various espresso drinks from 5:30 AM to 11:30 PM, fold 50 different novel items of clothing in a new home, and assemble and label 59 boxes for chocolate packaging in a real factory.

And π*0.6 achieved a success rate exceeding 90% across these three core tasks. Particularly in coffee-making, the improvement was significant, with both throughput and success rates doubling compared to the base model.

Currently, training for robot foundation models relies heavily on manually collected demonstration data, which is labor-intensive and limits performance ceilings. The Recap method opens new possibilities by making robots' autonomous practice data one of the core sources for training.

In fact, the breakthrough of π*0.6 lies not only in improved task performance but also in freeing robots from over-reliance on human demonstration data, establishing autonomous practice data as the core of training.

As robots are deployed at scale in real-world scenarios in the future, massive amounts of autonomous practice data will form a data flywheel: broader deployment leads to more diverse data, accelerating model performance improvements; enhanced performance, in turn, drives wider deployment, ultimately enabling robots to surpass human operational limitations and achieve superhuman robustness and efficiency.