- An arXiv paper published October 9, 2026 introduces TGIT, a front-end instruction translator that converts vague human commands into precise, machine-executable steps for frozen aerial Vision-Language Navigation agents, without modifying the underlying model.
- TGIT brings frozen OpenFly navigators to a 37.93% success rate on weak, high-level instructions and transfers zero-shot to real human commands, with gains also recorded on the held-out CityNav and AirVLN benchmarks.
Aerial drones can thread through complex three-dimensional environments with increasing confidence, but ask one to “fly around the building and check the roof” and it will often fail, not because the navigation model is weak, but because it was never trained to parse instructions that vague. A paper posted to arXiv on October 9, 2026, “Speaking the Navigator’s Language: Trajectory-Grounded Instruction Translation for Frozen Aerial VLN Agents,” proposes a targeted fix: a translation layer called TGIT that rewrites human language into the precise, trajectory-aligned commands frozen aerial agents actually respond to, leaving the underlying model untouched.
The Command Gap
Vision-Language Navigation agents are trained to follow detailed, step-by-step directional instructions. Human operators don’t talk that way. The result is a mismatch that hits hardest with “frozen” navigators, models whose weights are fixed after pre-training to preserve general capabilities and avoid the cost of retraining. Frozen agents are practically attractive, but their inflexibility makes them especially brittle when instructions drift toward natural speech. TGIT targets that specific failure mode rather than trying to retrain its way out of it.
How the Translation Works
TGIT sits in front of the frozen VLN agent as a modular pre-processor. It takes a high-level, underspecified instruction and rewrites it into the detailed, trajectory-aligned form the downstream agent expects. The grounding mechanism is the key move: TGIT uses information from predicted flight paths to infer which precise actions would fulfill the original intent, then produces instructions in the vocabulary the frozen navigator already understands. The system learns this mapping through iterative refinement rather than supervised labeling, which means it does not require a large paired dataset of human instructions and agent-ready commands. The frozen agent beneath it never changes.
Benchmark Results
On frozen OpenFly navigators given weak, high-level instructions, TGIT reached a 37.93% success rate, a figure the paper presents as a meaningful recovery from the near-floor performance such agents typically show on natural language inputs. The system also transferred zero-shot to real human commands it had not seen during training. On held-out benchmarks including CityNav and AirVLN, TGIT either improved or restored navigation performance, suggesting the approach generalises across urban and open aerial environments rather than fitting tightly to a single test distribution. For context on how arXiv-stage results in aerial robotics have held up under scrutiny, the discrete tokenization challenges facing vision-language-action models offer a useful parallel in related work on VLA navigation limits.
Why Frozen Models Matter
Retraining a large foundation model is expensive, slow and risks what researchers call catastrophic forgetting, where the model loses general skills as it adapts to a narrower task. Keeping the base agent frozen avoids all of that. TGIT’s modular design means the translation layer can be swapped or updated without touching the navigation model itself, which shortens the adaptation cycle considerably for teams deploying drone fleets across varied operational contexts. The practical implication is that an operator could reconfigure how agents interpret instructions without scheduling a retraining run. Whether that flexibility holds across instruction styles significantly more varied than the paper’s test sets is something the arXiv preprint does not yet establish.


