Are the environments in Isaac Gym completely independent?

[330184646]

I trained a strategy for object grasping using reinforcement learning with 1000 parallel environments. I fixed all random seeds as much as possible. My strategy achieved a reproducible success rate of 88.5% during repeated runs. However, intriguingly, when I set the RL strategy’s first action to 0 in a specific environment (solely for testing purposes, without considering its practical significance), I observed that the success rate fluctuated by more than 0.1% (indicating that the action change in this environment affected other environments). Furthermore, through multiple repeated runs, I identified the 11.5% of environments that failed. When I set the actions of these environments to 0 while keeping the actions of the other environments unchanged, this should not have affected the overall success rate. However, surprisingly, it led to a significant drop in the success rate. I am wondering whether the environments in Isaac Gym are entirely independent or if the success or failure of one environment could influence others to some extent.