EmbodiedWorld-200K: Toward Embodied World Agents via Embodied-Planning Dataset and Interactive World Models

Key points

• Building embodied agents that can perceive, reason, and act across diverse 3D worlds is a central research direction toward artificial general intelligence.1
• We refer to such agents as Embodied World Agents (EWAs), whose defining property is the ability to understand diverse visual environments and execute general actions in open-world conditions.1
• To address this, we propose a novel training scheme that employs a video world model as a low-cost environment proxy and performs interactive reinforcement-learning (RL) post-training of the EWA within it.1
• Empirically, pretraining on EmbodiedWorld-200K substantially improves the open-world embodied planning ability of the EWA, and further RL post-training inside a world model yields clear gains on out-of-distribution scenes.1

Sources (1)

• [1]EmbodiedWorld-200K: Toward Embodied World Agents via Embodied-Planning Dataset and Interactive World Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 07:52 AM

  To advance EWA research, we study open-world embodied navigation-planning and release EmbodiedWorld-200K, a large-scale dataset of over 200K samples spanning diverse scenes.

  Building embodied agents that can perceive, reason, and act across diverse 3D worlds is a central research direction toward artificial general intelligence.

Extractive summary: sentences quoted from the sources.