Researchers introduce VICTR, a method that helps vision-language-action robots choose useful parts of a demonstration while carrying out a task. Instead of retrieving examples based only on visual similarity, it estimates how far the robot has progressed relative to a reference video, then supplies demonstration chunks from the matching stage to a π0.5 policy.
On unseen, out-of-domain real-world manipulation tasks, the team reports a 23.8% success rate, compared with 6.2% for a fine-tuned π0.5 baseline. The paper also reports that task progress rose from 28.9% to 55.5%; it does not establish how the method performs in broader deployment.
Why it matters
For robotics researchers, VICTR suggests that matching demonstrations by task progress may help robots handle unfamiliar manipulation tasks; performance beyond the reported tests remains unknown.
