Robo2VLM

arXiv

TL;DR

Motivations & Innovations

Data Recipe

Scene-Interaction Understanding

Embodied Scene Understanding:

Input:

• a task description in nature language
• images from different camera views

Output:

• using an off-the-shelf LLM such as Qwen 2.5 to obtain target object, scene, task and skill description.

Manipulation Phase Segmentation:

Input:

• 12

Output:

• 12