D-World: An Explicit World-Language-Action Model with Recursive Self-Improvement under Dual Control

Key points

• Frontier large language models (LLMs) now reason about space well enough to program robots, yet the world models behind robot policies are implicit, held in network weights or predicted video, and cannot be inspected or corrected directly.1
• A perception agent writes the scene as an executable digital twin using generated vision and geometry code.1
• A planning agent searches for actions in simulation and, after each executed action chunk, revises the twin's object states and named physical parameters against what the robot measured.1
• We model the choice of actions that both advance the task and expose model error as a dual control problem and show that the cost of incorporating evidence into the twin sets the threshold at which probing pays, so cheap local revision counteracts the turn-off of cautious controllers.1

Sources (1)

• [1]D-World: An Explicit World-Language-Action Model with Recursive Self-Improvement under Dual Control
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 05:05 PM

  We present D-World, an explicit world-language-action model built on the Sense, Plan and Act decomposition.

  Frontier large language models (LLMs) now reason about space well enough to program robots, yet the world models behind robot policies are implicit, held in network weights or predicted video, and cannot be inspected or corrected directly.

Extractive summary: sentences quoted from the sources.