my own env doesn't converge

Hey,

I added my own env dopamine ( repo: [https://github.com/KatyNTsachi/Hierarchical-RL](url) )
I tried to train DQN agent on it, but it won't converge.
This is reward graph
I did minor changes to the original code, I checked the returned picture and the returned reward and it seems like it's doing what it should do.
The game that I created is very easy, you just need to take prizes that shows on the screen.
the game looks like this:
the green ball is the player and the pink ball is the prize.

Any idea how to solve this problem?

Thanks,
Tsachi