Roboticsworld modelsStory 06
Imagined bag handling helps a real robot lift
What changedWorldContact researchers trained a model to predict contact with flexible shopping bags and generate extra robot-training data. A real arm's single-attempt lifting success rose from 65% to 95% in their comparison. The wider task set remains simulated, and the reported generation speed excludes rendering and saving files.
The useful part
Why it matters
A learned model of contact can generate more training experience for flexible objects that are difficult to simulate quickly.
Keep in mind
Good to know
The tenfold speed comparison excludes rendering and disk I/O. Real-robot validation concerns a narrow bag-lifting task, not all 16 simulated tasks.
Evidence