Issue: Inconsistencies between image input selection and action representation

[LKlkun]

First, thank you for releasing Nav-R1 and making both the code and models publicly available — this is a very exciting and valuable project.I have two questions that I’d appreciate some clarification on:

Question 1 — Image Input Logic:

In the paper, it is stated that the model receives five past egocentric observations (Images 1–5) and one current egocentric observation (Image 6).

However, in the current implementation of build_cot.py, the following code is used:

selected_images = images[:max_images] if max_images else images

This simply slices the list of skybox images without distinguishing between past and current viewpoints.
Question: How should we correctly provide the “current timestep” image (Image 6) to match the setup described in the paper?

Question 2 — Action Format

The current CoT output represents continuous 3D navigation actions with coordinates and yaw angles, for example:

{"action": "MOVE", "from": [...], "to": [...], "heading_yaw_deg": -56.03}

Meanwhile, the paper describes discrete symbolic actions, such as:

Move forward 25 cm, Turn left 30°, or Stop.

Question: Should the CoT generator (build_cot.py) quantize these continuous movements into discrete symbolic tokens before saving, or are the continuous values intentionally used for both training and evaluation?

Thank you for your time and for this excellent work. I look forward to your clarification.

[waynechu1021]

First, thank you for releasing Nav-R1 and making both the code and models publicly available — this is a very exciting and valuable project.I have two questions that I’d appreciate some clarification on:

Question 1 — Image Input Logic:

In the paper, it is stated that the model receives five past egocentric observations (Images 1–5) and one current egocentric observation (Image 6).

However, in the current implementation of build_cot.py, the following code is used:

selected_images = images[:max_images] if max_images else images

This simply slices the list of skybox images without distinguishing between past and current viewpoints. Question: How should we correctly provide the “current timestep” image (Image 6) to match the setup described in the paper?

Question 2 — Action Format

The current CoT output represents continuous 3D navigation actions with coordinates and yaw angles, for example:

{"action": "MOVE", "from": [...], "to": [...], "heading_yaw_deg": -56.03}

Meanwhile, the paper describes discrete symbolic actions, such as:

Move forward 25 cm, Turn left 30°, or Stop.

Question: Should the CoT generator (build_cot.py) quantize these continuous movements into discrete symbolic tokens before saving, or are the continuous values intentionally used for both training and evaluation?

Thank you for your time and for this excellent work. I look forward to your clarification.

I also have the same question. It seems that they did not collect images at different time steps in the simulator; directly using images from Skybox is incorrect, and the code implementation doesn't match the paper in many ways. I noticed similar errors in their newly released MobileVLA-R1, and I have doubts about the performance of their implementation.

[inf-usb]

I also have the same question. It seems that they did not collect images at different time steps in the simulator; directly using images from Skybox is incorrect, and the code implementation doesn't match the paper in many ways. I noticed similar errors in their newly released MobileVLA-R1, and I have doubts about the performance of their implementation.

I have the same question. I don't quite understand why they directly use images from Skybox.