Abstract:

We address key limitations in existing datasets and models for task-oriented hand-object interaction video generation, a critical approach of generating video demonstrations for robotic imitation learning. Current datasets, such as Ego4D, often suffer from inconsistent view perspectives and misalign...

Key points:

• Research on large language models
• Engineering application

Source: TASTE-Rob: Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation