[DOC] How to generate nav_text for custom data inference?

[summer4272]

When using Alpamayo 1.5 for inference on custom data, the model accepts an optional nav_text input (e.g., "Turn left in 30m") via helper.create_message(nav_text=...) to condition trajectory predictions.

However, the PhysicalAI-AV public dataset does not include any navigation field, and there is no code in the released repository that generates nav_text. The only reference is the 20 pre-computed samples in nav_demo_samples.json with format "Turn {left/right} in {distance}m", which include nav_maneuver and distance_m fields.

My questions:
How is nav_text generated for training or inference?

For users running inference on their own driving data, what is the recommended approach to generate nav_text?

Compute from future trajectory (detect heading change > threshold, measure along-path distance)?

Or is it acceptable to just provide simple commands like "Turn left" without distance?

Any guidance on the expected format and generation method would be very helpful for the community.

[super-anova]

Hi, thanks for reaching out!

For nav_text, our recommended way to generate them is to use the future ego trajectory's GPS data as reference, and query a map or routing service to get the turn by turn nav instructions.

That said, physical AI AV data do not have GPS data so you won't be able to do this for them. For your own data, if they have GPS data, then it may be feasible to do this.

In terms of instruction format, we trained the model with the distance information in nav_text and we have observed that the model attends to the provided distances in generating predicted trajectories. So simply provide "Turn left" may be slightly out of distribution, but might still work to some extent. Generally in our data navigation instructions look like "Turn left in xx m", "Merge slight right in xx m", "Continue in xx m", etc.