TGIT Translates Short Drone Commands and Lifts Navigation Success from 11% to 33%

Aerial vision-and-language navigation (VLN) models are trained on detailed commands aligned to the flight trajectory. People talk differently: briefly, and focused on the goal. On a frozen OpenFly navigator, this gap cuts the mission success rate (SR) from 31.03% to 11.33%. The authors of a new arXiv paper propose the Trajectory-Grounded Instruction Translator (TGIT). It rewrites a short command into a form the agent understands, and it does not require fine-tuning the navigator itself. On real human instructions, it raises SR from 11.33% to 32.51%.

Where the human-drone link breaks

The navigator sees the world through its training set. There, instructions read like a step-by-step route: fly along the building, turn at the intersection, descend by the red roof. A user will say, "fly to that tall building by the river." For a person, that is enough. For an agent used to detail, it is almost noise.

The paper measures this gap directly. On detailed original commands, the frozen OpenFly scores 31.03% SR. On short, intent-oriented instructions, the result drops to 11.33%. Retraining the navigator for every speaking style is expensive and inconvenient, especially if it is already deployed and frozen. So the authors changed the input rather than the model.

How TGIT works

Training a translator requires pairs of a short command and a detailed command. Real pairs are scarce, so the authors built them synthetically. They gave a language model several human-written examples of the style and asked it to turn the original commands into "weak" (Weak) ones: short and intent-focused. These Weak commands alone yield 15.27% SR. That is slightly better than live human instructions, but noticeably worse than the originals.

TGIT works as a front end placed before the navigator. It takes a Weak input and produces a command the agent can execute. The translator trains on trajectory outcomes: it observes where the frozen navigator actually flew after a given phrasing and tunes its translation to that behavior. Hence the name "trajectory-grounded."

Key figures from the abstract:

• frozen OpenFly on detailed commands: 31.03% SR;
• the same model on short human instructions: 11.33% SR;
• Weak commands without translation: 15.27% SR;
• Weak input through TGIT: 37.93% SR;
• real human instructions through TGIT, zero-shot: 32.51% SR;
• OpenFly held-out split: from 4.95% to 20.79%.

The 37.93% figure deserves a separate look. It is higher than the 31.03% on the original detailed commands. Apparently, the translator does not merely restore lost details; it learns to phrase commands the way the navigator handles best. Still, treat this with caution: the full paper is the place to compare these numbers against the experimental conditions.

Transfer matters too. The translator trained on synthetic Weak commands, yet worked on real people without fine-tuning: 11.33% became 32.51%. For the CityNav and AirVLN benchmarks, the authors report "recovery" of quality. The abstract gives no numerical values, so the scale of the effect there is still unknown.

What you can take into production

The main practical idea is that the adapter only needs to observe the behavior of another model. No access to its weights, gradients or training data is needed. Teams that run off-the-shelf navigation models for drones or robots could put such a layer in front of the system and leave untouched whatever is already certified or tested.

There are limitations. The results come from simulation benchmarks, not real flights. The Weak commands were generated by a model from a few samples, so the real-world variety of speech, accents and voice-input noise remains an open question. Before building a product, it makes sense to test TGIT on your own scenarios and your own query style. If the effect holds up, translating intent into an executable instruction could become a separate layer in the autonomous-agent stack, rather than a task that requires retraining every new model.