US firm teaching humanoid robot brains to do laundry make coffee light candles

Описание: US firm teaching humanoid robot brains to do laundry, make coffee, light candles
A technician uses two joysticks to teleoperate small robotic arms that fold T-shirts. Larger robots shuffle pantry goods between boxes.

Behind a nondescript metal door in the Mission District, marked only with the symbol “π,” a new robotics startup is pushing toward one of Silicon Valley’s most ambitious goals: giving machines the ability to learn physical tasks with human-like reliability.

The warehouse belongs to Physical Intelligence (PI), a company that blends robotics and advanced AI training methods to improve how machines handle real-world manipulation. Inside, workers and machines move rapidly around a crowded floor.

A technician uses two joysticks to teleoperate small robotic arms that fold T-shirts. Larger robots shuffle pantry goods between boxes. Another worker tests a wrist-mounted pincer fitted with a webcam. Unfinished robot parts cover nearly every surface.

Teaching humanoid robot brains

The young startup has quickly become one of the most closely watched companies in the field. Last week, PI announced it had raised $400 million from investors, including OpenAI and Amazon founder Jeff Bezos, valuing it above $2 billion.

While many AI systems excel in digital tasks, robotic learning still struggles with reliability. Most systems learn from human demonstrations, the equivalent of watching how a box is folded or how an espresso drink is made.

But that approach often breaks down when a robot deviates from the demonstration, causing small errors to compound until the task fails.

PI says it has developed a method intended to overcome that problem. The company calls the technique Recap, short for Reinforcement Learning with Experience and Corrections via Advantage-conditioned Policies .

It is designed to let robots learn the way humans do: by receiving instruction, being corrected, and then practicing on their own.

In recent tests using Recap, PI trained a new version of its vision-language-action model, known as π*0.6, to perform tasks such as folding laundry, assembling shipping boxes, and making espresso drinks.

According to the company, performance on some tasks more than doubled, with failure rates dropping by more than half. The system was able to make coffee continuously for a day, fold clothing for hours in a home environment, and assemble packaging boxes at factory-relevant speeds.

Researchers say the reliability gap, the difference between partial success and near-perfect execution, is one of the main barriers preventing robotics from operating at scale in warehouses, kitchens, and manufacturing floors.

The difficulty stems from physical interaction: when a robotic gripper misses its target by even a few millimeters, the resulting misalignment creates a situation the robot has never seen in its training data.

Doing laundry, making coffee, lighting candles

Recap tries to correct this by introducing two additional data streams beyond demonstrations. First, human operators step in when a robot begins to fail, providing corrective actions that teach the system how to recover from mistakes.

Second, the robot evaluates its own performance through reinforcement learning, assigning credit or blame to the actions that led to success or failure.

The model uses a “value function” to determine which moves improved its position toward completing a task, even when the effect is only visible later.

PI says this process allows robots to learn from imperfect experiences rather than discarding them, giving the system access to far more training data than manual demonstrations alone could provide.

The company argues that as robots are deployed more widely, this kind of autonomous learning will be essential to scaling AI-enabled labor.

For now, PI trains its models on a range of household and industrial tasks that require dexterity, timing, and object understanding. Laundry folding demands generalization across different fabrics and shapes. Box assembly requires precise sequencing and repeated execution.

Espresso making, one of the most complex tasks, combines robotic manipulation with long-horizon actions such as grinding coffee, operating the machine, and cleaning equipment.

The results, according to the company , show the potential of combining human guidance with machine practice, a hybrid approach that researchers have long theorized but struggled to implement at scale.

Physical Intelligence says it plans to expand partnerships with companies that operate commercial robots and are seeking greater autonomy.

The firm has not said when its models will be deployed outside controlled environments. Still, it argues that robots capable of learning from experience could eventually achieve performance levels exceeding those of human workers on repetitive physical tasks.