TGIT
A Trajectory-Grounded Instruction Translator placed upstream of a navigation model. It uses trajectory-level reinforcement learning to convert brief, intent-style user expressions into detailed instructions understandable by the navigator, without modifying the underlying navigation model.
Timeline (last 90 days)
TGIT (Trajectory-Grounded Instruction Translator) is a translation layer that keeps the navigator completely frozen and converts a user's brief, intent-based weak instructions into the command format that the navigator can understand and execute.
TGIT employs trajectory-level rewards by feeding translated instructions into the frozen navigator for execution and using whether the flight trajectory reaches the target as a feedback signal to update the translator's parameters.
The paper uses manually written style examples to prompt a language model to convert original exhaustive commands into intent-centered weak instructions, batch-generating paired training data.
Using only synthetic weak instructions fed directly into the navigator, the success rate was 15.27%, falling between the exhaustive command rate of 31.03% and the real user instruction rate of 11.33%.
After TGIT training, the success rate for weak instruction input increased from 15.27% to 37.93%, surpassing the original exhaustive command rate of 31.03%.
TGIT's zero-shot transfer to real human instructions saw the success rate jump from 11.33% to 32.51%.
On an unseen OpenFly test set, TGIT improved the success rate from 4.95% to 20.79%.
TGIT also achieved performance recovery on two independent benchmarks, CityNav and AirVLN.
TGIT proposes a decoupled paradigm that separates the tasks of understanding user intent and executing navigation, optimizing each in the most appropriate manner.
TGIT fully externalizes adaptation to the input side with zero intrusion into the underlying model, allowing integration regardless of whether the navigator is an open-source weight or a commercial API.
3 more timeline events
All Facts (13)
TGIT (Trajectory-Grounded Instruction Translator) is a translation layer that keeps the navigator completely frozen and converts a user's brief, intent-based weak instructions into the command format that the navigator can understand and execute.
50%UnverifiedTGIT employs trajectory-level rewards by feeding translated instructions into the frozen navigator for execution and using whether the flight trajectory reaches the target as a feedback signal to update the translator's parameters.
50%UnverifiedThe paper uses manually written style examples to prompt a language model to convert original exhaustive commands into intent-centered weak instructions, batch-generating paired training data.
50%UnverifiedUsing only synthetic weak instructions fed directly into the navigator, the success rate was 15.27%, falling between the exhaustive command rate of 31.03% and the real user instruction rate of 11.33%.
50%UnverifiedAfter TGIT training, the success rate for weak instruction input increased from 15.27% to 37.93%, surpassing the original exhaustive command rate of 31.03%.
50%UnverifiedTGIT's zero-shot transfer to real human instructions saw the success rate jump from 11.33% to 32.51%.
50%UnverifiedOn an unseen OpenFly test set, TGIT improved the success rate from 4.95% to 20.79%.
50%UnverifiedTGIT also achieved performance recovery on two independent benchmarks, CityNav and AirVLN.
50%UnverifiedTGIT proposes a decoupled paradigm that separates the tasks of understanding user intent and executing navigation, optimizing each in the most appropriate manner.
50%UnverifiedTGIT fully externalizes adaptation to the input side with zero intrusion into the underlying model, allowing integration regardless of whether the navigator is an open-source weight or a commercial API.
50%UnverifiedTGIT cannot correct the capability boundaries of the navigator itself; if the navigator is inherently weak in a certain type of scenario, translation will be of no avail.
50%UnverifiedTGIT's trajectory-feedback training depends on the availability of a simulation environment, and the cost of obtaining training signals will increase significantly in purely real-world physical deployment scenarios.
50%UnverifiedThe TGIT-related work was published as a newly released arXiv preprint.
50%