Robots Learn to Fold Pants: The Future of AI?

Physical Intelligence, a San Francisco-based company, is developing general-purpose robots through a continuous data collection and training loop, akin to an "AI for robots." Researchers are training foundation models that are then evaluated on stations performing tasks like folding laundry and peeling vegetables. This approach uses accessible hardware, costing around $3,500 per arm, emphasizing that strong AI intelligence can compensate for less advanced hardware.

The company, co-founded by UC Berkeley professor Sergey Levine, has secured over $1 billion in funding. Unlike competitors such as Skild AI, which focuses on immediate commercial deployment and revenue generation, Physical Intelligence prioritizes fundamental research. This strategy involves cross-embodiment learning and diverse data sources, aiming to lower the cost of adapting autonomy to new robot platforms.

Physical Intelligence is currently testing its systems with select companies in logistics, grocery, and manufacturing. While some tasks are deemed ready for real-world automation, the company’s long-term vision, without a defined commercialization timeline for investors, centers on achieving superior general intelligence. This approach contrasts with rivals who believe commercial deployment creates a data flywheel for faster model improvement.

The development of general-purpose robotic intelligence is a rapidly advancing field, with companies like Physical Intelligence and Skild AI pursuing different paths. Physical Intelligence's strategy relies on extensive research and development, funded by significant investment, to build a foundation for future robotic applications. Hardware development, however, remains a significant challenge due to its inherent complexities and slower iteration cycles compared to software.