RESEARCH · RESEARCH · #1777
TGIT: trajectory-grounded translator improves frozen OpenFly aerial VLN on short instructions
This arXiv preprint (arXiv:2610.10635v1) identifies an "instruction gap" where short, intent-driven user instructions drop a frozen OpenFly aerial VLN agent's success rate from 31.03% to 11.33%. The authors first use a prompted language model to generate paired intent-centered "Weak" commands (raising SR to 15.27%), then introduce the Trajectory-Grounded Instruction Translator (TGIT), a front-end trained from the navigator's trajectory outcomes that boosts Weak-input SR to 37.93%, transfers zero-shot to real human instructions (11.33% -> 32.51%), improves held-out OpenFly (4.95% -> 20.79%), and yields recovery on CityNav and AirVLN.
KEY POINTS
- This arXiv preprint (arXiv:2610.10635v1) identifies an "instruction gap" where short, intent-driven user instructions drop a frozen OpenFly aerial VLN agent's success rate from 31.03% to 11.33%.
- The authors first use a prompted language model to generate paired intent-centered "Weak" commands (raising SR to 15.27%), then introduce the Trajectory-Grounded Instruction Translator (TGIT), a front-end trained from the navigator's trajectory outcomes that boosts Weak-input SR to 37.93%, transfers zero-shot to real human instructions (11.33% -> 32.51%), improves held-out OpenFly (4.95% -> 20.79%), and yields recovery on CityNav and AirVLN.
- TGIT provides a practical front-end approach that lets existing frozen aerial VLN navigators accept terse, human-style instructions and substantially recover navigation success without retraining the navigator.
WHY IT MATTERS
TGIT provides a practical front-end approach that lets existing frozen aerial VLN navigators accept terse, human-style instructions and substantially recover navigation success without retraining the navigator.