search

robodhruv / drive-any-robot

Official code and checkpoint release for "GNM: A General Navigation Model to Drive Any Robot".
MIT License
261 stars 41 forks source link

rosbag file for deployment #8

Closed james-yoo closed 1 year ago

james-yoo commented 1 year ago

Hi @PrieureDeSion, Is it enough to prepare image and odometry topic in rosbag file? And is the accuracy of the odometry an important factor in graph generation?

robodhruv commented 1 year ago

Hi @james-yoo , Yes, that is all the information you should need in the rosbag! We did not find the odometry accuracy to be too important and used all available odometry in our public datasets out of the box -- this includes odometry from lidar, wheel encoders, VIO etc. If you have an estimate of the accuracy of odometry that is very poor, can you share it? My expectation is anything under 2-5% should not be a problem at all.

james-yoo commented 1 year ago

Hi @PrieureDeSion, Sometimes it will be needed to travel hundreds of meters to kilometers for logging data. So, the accumulated distance error will be very large, which will cause problems in generating the graph. In this case, do I have to use precise positioning system (e.g. RTK, LiDAR SLAM, etc) to correct bias in odometry?

robodhruv commented 1 year ago

I don't think that should be too big of a problem because the odometry estimates are only used to obtain distance labels on the order of 20-40 time steps (say 10 seconds into the future). This relates closely to how the high-level planning is set up -- we only use nearby pairs of states to label positives in training and train the policy on these positives. I don't think long-range drifts will be a problem as long as you have relatively good local odometry.

james-yoo commented 1 year ago

Thanks for your detailed explanation.