[Question]: Frame of the output trajectory

[liujy25]

Question

Thanks for your work. I am working on real robot experiments. I noticed there are some transform parameters in realworld agent functions, such as pose and intrinsic, but they are not used during inference time. Are these parameters set only for training? When predicting the trajectory with real robots, the model receive egocentric rgbd data and output trajectory in robot base frame directly?

[yuqiang-yang]

• Pose and intrinsic are NOT used for both training and inference.
• The output trajectories are in robot base frame, following ROS convention.

[liujy25]

Thanks for your reply. I ran into some issues during real-robot deployment. In relatively narrow corridors, the goal point generated by S2 looks fine, but the trajectory generated by S1 shows a noticeable lateral drift.

My deployment platform:

i9-14900KF + RTX 4090 D

The robot camera is a ZED2, mounted at a height of about 1.3 m with a downward pitch angle of roughly 10 degrees. The output image resolution is 960×540.

I used the HTTP server and client code from the repository; the only modification was changing the depth encoding to 32FC1.

Below are the detailed visualization results. The generated trajectory is approximately 2–3 m long.

[liujy25]

I noticed that this implementation does not use the critic head in NavDP. Is there a reason it was removed or disabled? Also, does the provided checkpoint include trained weights for the NavDP critic head? When I run it, the critic values for different candidate trajectories are almost the same: