The headquarters of Physical Intelligence in San Francisco features a unique pi symbol marking its entrance, but once inside, the atmosphere is bustling with activity. The interior resembles a vast concrete expanse, furnished with blonde-wood tables scattered with items from lunch to robotics equipment. Various tables are cluttered with monitors and spare parts, showcasing robotic arms engaged in tasks that range from folding pants to peeling zucchinis.
During a visit, one robotic arm struggled to turn a shirt inside out, while another successfully peeled a zucchini, demonstrating its capability. Sergey Levine, co-founder of the company and an associate professor at UC Berkeley, likened the operations to “ChatGPT for robots.” He described the ongoing testing phase where data collected from various robot stations, including homes and warehouses, is utilized to train general-purpose robotic models.
This continuous loop of data collection and model evaluation is critical to the company’s mission, as they seek to teach robots the fundamental motions necessary to handle a variety of tasks. Physical Intelligence also maintains a test kitchen to expose robots to different environments, utilizing standard hardware to further their research and development efforts.