Physical Intelligence, a hot robotics startup, says its new robot brain can figure out tasks it was never taught

Breakthrough in Robotic Generalization

Physical Intelligence, a San Francisco-based robotics company established just two years ago, has released research showing its newest AI system, π0.7, gives robots the ability to tackle unfamiliar tasks. Rather than relying on specific task training, the new model demonstrates compositional generalization: it combines skills learned in various settings to solve problems it hasn't directly encountered.

How π0.7 Stands Out

Traditionally, robots are trained for fixed tasks via direct data collection and specialized models. π0.7 overturns this by remixing knowledge from different experiences, allowing robots to handle novel situations with minimal guidance. In one clear example, π0.7 successfully operated an air fryer after only minimal prior exposure, synthesizing data from seemingly unrelated episodes. When guided through verbal instructions similar to onboarding a human worker, the robot completed tasks effectively, such as cooking a sweet potato.

Implications & Limitations

This ability could let robots adapt quickly in new environments and improve on the fly without major additional training. The team, however, admits there are still many boundaries. The model can’t handle multi-step tasks from a single command but improves greatly with step-by-step coaching. Results also depend on how humans phrase instructions—a quick adjustment in prompt wording can dramatically boost success rates.

Investor Excitement and the Path Forward

Physical Intelligence’s progress has attracted major investment, with over $1 billion raised and a valuation topping $5.6 billion, reportedly targeting even higher numbers. The blend of Silicon Valley investment experience among company founders and remarkable research advances helps explain this momentum. While generalization in robotics remains less flashy than in language AI, it’s a foundational step for practical, flexible machines.

When Will Generalist Robots Be Common?

The company remains cautious about predicting when such generalist robots might see widespread deployment but is optimistic about continued rapid progress. As benchmarks and standards evolve, real-world applications are likely to follow in the coming years.

Read the original article on TechCrunch for more details.