US firm teaching humanoid robot brains to do laundry, make coffee, light candles
SMRTR summary
Behind a metal door marked only with the symbol "π" in San Francisco's Mission District, technicians guide robotic arms through the delicate art of folding T-shirts while larger machines shuffle pantry goods between boxes. This is Physical Intelligence, a startup that just raised $400 million from investors including OpenAI and Jeff Bezos, betting they can solve robotics' reliability problem.
The company has developed a technique called Recap that teaches robots to learn like humans do: through instruction, correction, and practice. Unlike traditional systems that crumble when robots deviate even slightly from demonstrations, this approach lets machines recover from mistakes and improve through experience.
Recent tests showed dramatic improvements. Robots using their latest model folded laundry for hours, made coffee continuously for an entire day, and assembled packaging boxes at factory speeds. Performance on some tasks more than doubled while failure rates dropped by half.
The breakthrough addresses what researchers call the reliability gap, where missing a target by mere millimeters can cause complete task failure. Physical Intelligence believes this hybrid of human guidance and machine practice could finally push robotics toward truly autonomous operation in warehouses, kitchens, and manufacturing floors across the country.
SMRTR provides this summary for quick context. The original article belongs to Interesting Engineering.
Read the original article