The AI Model You Use Every Day Is Getting Better at Controlling Robots. Claude Fable 5.2 More Than Doubled Its Predecessor's Score on a Simulated Robot

When the next Claude or GPT model ships, I want to see an old robot try again. Can it now recover from a mistake that used to require a human?

TL;DR

  • AI models are being trained to control physical robots, with examples like Astra controlling a kitchen arm to prepare a dinner party.
  • The performance of AI models in robotics is being measured against theories like Rich Sutton's 'Bitter Lesson', which suggests general AI methods will eventually surpass specialized ones.
  • New versions of AI models like Claude are showing significant improvements in simulated robot tasks, doubling their predecessors' scores.
  • The article explores the historical context of factory robots designed for fixed environments and the cost of teaching robots new tasks.
  • Recent tests show general AI models succeeding in robot control, including simulated arms and real-world trials, with some agents writing their own routines.
  • Specialized AI in robotics is also accelerating due to factors like open-source collaboration and AI coding assistants.
  • Future developments in 2027 will focus on testing these AI models in real kitchens to see if they can overcome errors that previously required human intervention.