TGIT Translates Weak Instructions to Improve Aerial VLN Success
October 9, 2026
The Trajectory-Grounded Instruction Translator (TGIT) enables frozen aerial vision-and-language navigation agents to handle short, intent-driven user commands. By learning from trajectory outcomes, the front-end raised zero-shot success rates on real human instructions from 11.33% to 32.51%.
HOW THIS AFFECTS YOU
●
builderYou can deploy this to bridge the gap between vague user intents and precise robot command requirements.
●
researcherThis demonstrates how trajectory-grounded training can mitigate the instruction gap in frozen agents.