Researchers led by Ruizhe Xie have presented VLN-Flash, a decoder that speeds up vision-and-language navigation, in which a robot follows instructions given in words, according to a paper on arXiv. When such a policy generates its actions one after another, the robot has to wait for a chain of predictions that depend on each other, the authors write.
Because the set of possible actions is small, VLN-Flash scores every short plan in a single pass without extra training, and uses how far the best plan leads the next best to decide how many actions to carry out before looking again. It decoded actions almost three times faster than the usual step-by-step method on the Uni-NaVid and StreamVLN policies and raised success rates on the R2R-CE and RxR-CE benchmarks by up to 2.7 percentage points.
On a Unitree Go2 robot dog, it improved navigation success and cut the average wait between actions by about 160 milliseconds, not counting network transmission time.



