papersTODAY 04:00 UTC
DART-VLN tackles memory decay and looping in vision-language navigation agents
A new arXiv paper proposes a method called DART-VLN for memory-based agents that navigate environments using discrete vision-language instructions. It targets two failure modes seen at inference time: reliance on outdated stored information and agents getting stuck in repetitive behavior. The approach combines test-time memory decay with an anti-loop regularization technique to improve navigation robustness.