[Question] Multi-task RL with tasks having different observation terms

[meenalparakh]

Hello! Is it possible to set up multi-task RL training (where the tasks have different numbers of objects and corresponding observation terms) using Isaac-lab and get any insights on how to approach it? For example, I want to train a single PPO policy on five such tasks. How can I collect data across multiple tasks efficiently?

• Will it be very costly to spawn, say, 100 environments for a task and change the task after every few episodes?
• What about spawning environments for all tasks side-by-side, in which case there is no need to create scenes from scratch, but then how can I initialize it (e.g., 100 envs for task#1, 100 envs for task#2 ... in the same world) and obtain the different sets of observation and actions associated with them?
• Any other ways or resources to help with this?
  Thank you!

[XintongHe]

Hi, did you solve your problems? I have the same problem. Thank you