The GPT Moment for Robotics Is Here

Kwan Wong, co-founder of Physical Intelligence (Pi), still can't quite believe what his team has accomplished. "It still blows my mind to see a robot actually folding laundry," he says, "because I remember until basically until Chat GPT, I didn't know if this would exist even in my entire lifetime." Pi's mission is ambitious: build a model that can control any robot to do any task it is physically capable of, at a level of performance useful to people in all walks of life. Two years into the company, they are already deploying systems in real laundromats and e-commerce warehouses, and the pace of progress has been "pleasantly much faster than we expected."

Why Robotics Has Been So Hard

The robotics problem breaks down into three pillars: semantics, planning, and control. Semantics — understanding what objects are and what commands mean — got a major unlock from language models. Planning — figuring out the sequence of steps to accomplish a task — also benefited from those advances. But control — executing actions in real time in an environment that constantly changes — remained stubbornly difficult.