1 paper
Haki Darwish, Xiangyu Yin, Changwen Li +3
Benchmarks expose vision-language-action (VLA) policies to few canonical instructions, while exhaustive deployment testing is impossible. We introduce Object-Grounded Attention Mon…