Video2Policy is a breakthrough framework that transforms internet RGB videos into realistic simulation tasks for training robotic manipulation policies at scale.

• Solves the challenge of hallucinated tasks from LLMs and avoids complex digital twin alignment
• Reconstructs 3D scenes from 2D videos using neural scene understanding
• Enables diverse, realistic training without expensive physical robotics setups
• Demonstrates superior performance across manipulation tasks compared to baseline methods

This research significantly advances factory automation and robotics engineering by making policy training more accessible, diverse, and cost-effective for real-world applications.

Video2Policy: Scaling up Manipulation Tasks in Simulation through Internet Videos