search

leggedrobotics / viplanner

ViPlanner: Visual Semantic Imperative Learning for Local Navigation
https://leggedrobotics.github.io/viplanner.github.io/
Other
265 stars 27 forks source link

env_nums > 1 #20

Closed hutslib closed 2 months ago

hutslib commented 2 months ago

is it possible to set env_nums > 1to use multiple environment to training?

pascal-roth commented 2 months ago

Hi,

I assume you mean the num_envs passed to the TerrainSceneCfg. This is just done for the demo scripts. The training itself is not performed online, s.t. there is no notion of multiple environments for training. It uses offline collected data and runs batched, with a batch size depending on the size of your GPU.

Please let us know if that answers your question.

hutslib commented 2 months ago

Hi, clear explanation of the pipeline. Also, when the data is collected, is num_envs set as 1? The question is generated based on my trial of using isaaclab. I load your environment (e.g. a Carla town), and set num_envs > 1. The problem is the image collected by the robot can see robots in other environments.

pascal-roth commented 2 months ago

For the data collection, I am using a pipeline that just spawns a camera at different points in the space without actually using a robot. Consequently, many can be rendered at the same time without any robots in the image.

hutslib commented 2 months ago

Thanks for your clear explanation.

pascal-roth commented 2 months ago

No worries, let me know if you have any further questions.

Will close this issue for now.