[Question]: When performance falls short of expectations during reproduction on Go2

[Sung-Jae-Seong]

Question

Are there possible factors to consider when the performance during reproduction does not meet expectations?
Currently, I suspect the following variables may be affecting the results:

• Camera model (using RealSense D435i)
• System latency
• Motion blur in input images
• Camera height (default mounting position)
• Checkpoint ("InternVLA-N1-DualVLN")
• Differences in texture and lighting conditions

Among these (or other possible factors), which ones are known to have the most significant impact on performance?

Additionally, are there any critical setup details (e.g., camera calibration, frame rate, resolution, synchronization, etc.) that are easy to overlook but important for successful reproduction?

[yuqiang-yang]

1. The camera height, tile and motion blur have great influence on the performance.
2. System latency lower than 1s is acceptable if you do the right spatial-temporal alignment for the trajectory.
3. The model needs no camera calibration.
4. InternVLA-N1-DualVLN model is better than the InternVLA-N1-NavDP

[Sung-Jae-Seong]

Thank you for your response!
What does “tile” mean in this context?
Can it be controlled in the code (e.g., image tile size or crop position)?

[yuqiang-yang]

Tilt angle.