Physical Intelligence, a London-based robotics startup, has developed an AI brain capable of teaching humanoid robots new physical skills in days [1].

This development represents a significant shift in how machines learn to interact with the physical world. By reducing the time required to acquire new tasks, the company aims to make humanoid robots viable for high-value labor in factories and warehouses worldwide [2].

The AI model, known as π0.7 [3], is designed to accelerate the process of skill acquisition. Previously, training a robot to perform complex physical movements could take months of data collection and refinement [1]. The π0.7 model allows these systems to figure out tasks they were never explicitly taught, shortening the deployment cycle [3].

Physical Intelligence is targeting industrial environments where flexibility is critical. In traditional automation, robots are programmed for a single, repetitive motion. The new AI brain allows for a more general-purpose approach, meaning a robot could potentially switch between different warehouse duties without requiring a complete software overhaul [2].

While other firms in the sector are pursuing similar goals, the London startup focuses on the speed of learning. The ability to compress months of training into a few days could lower the barrier to entry for companies adopting robotics in their supply chains [1].

The company intends to deploy this technology globally to assist in labor-intensive sectors [2]. By optimizing the "brain" of the humanoid, Physical Intelligence seeks to bridge the gap between digital intelligence and physical dexterity [2].

“The AI brain can teach new physical skills in days rather than months”

The transition from month-long training cycles to day-long cycles suggests a move toward 'general-purpose' robotics. If AI models like π0.7 can successfully generalize physical tasks, the economic viability of humanoid robots increases because the cost of programming and downtime for new tasks decreases significantly.