What is it about?
Robots can learn to pick up and move objects by watching demonstrations. But a robot trained from one position can struggle when it approaches the same task from a different angle. EgoDemoGen creates new training examples from existing demonstrations. Each example shows what the robot would see from a new position and pairs that view with the movements it would need to make. We tested this approach in simulation and on a real robot.
Featured Image
Photo by Jason Leung on Unsplash
Why is it important?
When a robot starts in a different position, both what it sees and how it needs to move can change. By generating both parts of the demonstration together, EgoDemoGen could reduce the need to record the same task from many positions. Across five tasks on a real robot, average success from new viewpoints rose from 37% to 60% when we added the generated demonstrations to the original training data.
Perspectives
What I find most interesting is that a small change in where a robot stands can make a familiar task difficult again. I like to think about it from the robot’s point of view: What would I see from here, and how would I have to move to do the same thing? EgoDemoGen puts those two questions together in each training example.
Yuan Xu
Read the Original
This page is a summary of: EgoDemoGen: Egocentric Demonstration Generation for Viewpoint Generalization in Robotic Manipulation, November 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3767308.3835605.
You can read the full text:
Resources
Contributors
The following have contributed to this page







