EXPLORINEER
en

Archive — US — techcrunch

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

Physical Intelligence, a robotics startup, revealed that its latest model, π0.7, can perform tasks it hasn't been explicitly trained for by learning to synthesize skills from diverse contexts. This capability, termed compositional generalization, could lead to a significant leap in robotic AI, akin to advancements seen in language models, with the potential for real-time learning in new environments.

Key points

  • Physical Intelligence's model, π0.7, demonstrates compositional generalization, allowing it to synthesize knowledge from limited training data.
  • The model successfully operated an air fryer with minimal previous exposure, using step-by-step verbal instructions.
  • The researchers were surprised by the model's ability to generalize from disparate training episodes, indicating a major shift from rote memorization.
  • Concerns remain about the model's limitations, such as executing complex tasks from high-level commands and the lack of standardized robotics benchmarks for validation.
  • The team acknowledges that improvements in prompt engineering have a significant impact on outcomes, as seen in the air fryer experiment.
  • Physical Intelligence has raised over $1 billion in funds and has experienced a rapid increase in valuation due to investor enthusiasm.

Why it matters

  • The advancement of π0.7 may be part of a broader trend in robotics and AI toward models that can generalize and learn on the fly, similar to trends seen in ML with language models.
  • Historically, AI developed through specialized models; this represents a potential shift towards versatile systems that adapt to new contexts.
  • General-purpose AI capabilities are gaining traction, reflecting ongoing interest and investment in tech solutions that marry generalization with real-time adaptability.
  • The interest in generalization in robotics parallels developments in areas like autonomous vehicles and smart home technologies.

Physical Intelligence's π0.7 marks a potential breakthrough in robotics by facilitating a degree of generalization unseen in earlier models, paving the way for future advancements in AI and robotics, though challenges in validation and practical deployment remain.