arXiv Introduces TGIT: Boosting Aerial Vision-and-Language Navigation Success Rate to 37.93%
Summary:arXiv has introduced a novel method called the Trajectory-Grounded Instruction Translator (TGIT) for aerial vision-and-language navigation (VLN) agents. TGIT translates user-intent-driven, short instructions into detailed, executable commands by learning from trajectory outcomes, significantly improving navigation success rates. On the OpenFly navigator, TGIT boosts the success rate of weak inputs from 11.33% to 37.93% and successfully transfers to real human instructions (11.33% to 32.51%). The
The Challenge of Instruction Processing in Drone Navigation
Aerial vision-and-language navigation (VLN) systems typically rely on detailed, trajectory-aligned commands for training. However, the instructions provided by users in real-world applications are often short and intent-driven. This instruction gap leads to a significant drop in navigation success rates. For example, on the OpenFly navigator, the success rate decreases from 31.03% to 11.33%.
TGIT: A New Approach to Bridging the Instruction Gap
To address this issue, researchers have proposed the Trajectory-Grounded Instruction Translator (TGIT). TGIT enhances navigation success rates through the following steps:
- Instruction Translation: It uses a language model to convert original commands into weak commands that better align with user intent.
- Learning from Trajectory Outcomes: TGIT learns from the trajectory outcomes of the navigator to translate weak inputs into detailed, executable commands.
- Improving Success Rates: On the OpenFly navigator, TGIT boosts the success rate of weak inputs from 11.33% to 37.93% and successfully transfers to real human instructions (11.33% to 32.51%).
Technical Highlights
- Keeping the Navigator Frozen: TGIT acts as a front-end system and does not require retraining the navigator, reducing system complexity.
- Zero-Shot Transfer: TGIT can transfer its learned capabilities to real human instructions, demonstrating strong generalization capabilities.
- Cross-Task Applicability: TGIT performs well in both CityNav and AirVLN tasks, showcasing its applicability in different navigation scenarios.
Industry Impact and Developer Recommendations
TGIT offers a new method for instruction processing in drone navigation systems, effectively improving navigation success rates, particularly in complex environments. Developers can consider the following recommendations:
- Combine with Multimodal Data: Integrate visual and language data with TGIT to further enhance navigation accuracy.
- Optimize the Instruction Translation Model: Introduce more powerful language models or optimize the instruction translation algorithm to improve translation efficiency.
- Expand Application Scenarios: Apply TGIT to other types of navigation tasks, such as autonomous vehicles and robot navigation, to explore its broader application potential.
Conclusion
TGIT demonstrates the effectiveness of using instruction translation to improve navigation success rates in the drone navigation field, providing new ideas and technical pathways for future navigation system development.
— END —Source: ArXiv AI (cs.AI) (2026-10-10)
Tags: #Drones #Navigation Systems #TGIT #VLN #Instruction Translation
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments