Looking up old demonstrations scores 91.7% on a robot benchmark
Text
A Sereact/KIT study questions how much familiar robot tests reveal about generalization.
No new action-policy training, yet 91.7% success. A November 9 Sereact/KIT paper reports that result using demonstration retrieval on LIBERO.
The method combines actions from similar recorded situations. Vision-language-action models instead learn to generate actions.
With leading models already exceeding 95%, the comparison questions what high scores prove. Retrieval scored 82.4% after test conditions changed.
Pi0.5 remained stronger. These were simulation tests, and retrieval still used pretrained encoders. Training-free action selection is not a wholly untrained system.
Source submitted: arXiv, November 10, 2026, 03:44:58 Beijing time / November 9, 2026, 2:44:58 PM EST.
Share this article
What do you think of this story?
Anonymous feedback, up to 5 reactions per article.
Anonymous comments
Share your thoughts without signing in.
Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.
No comments yet. Be the first to share.