Consistent resetting through np.random.Generator - Githubissues

stepjam / RLBench

A large-scale benchmark and learning environment.

https://sites.google.com/corp/view/rlbench

Other

1.12k stars 228 forks source link

Consistent resetting through np.random.Generator #232

Open eugeneteoh opened 3 months ago

eugeneteoh commented 3 months ago

Currently, resetting is not consistent due to two things:

Sampled boundary is random and depends on global np seed (https://github.com/stepjam/RLBench/blob/master/rlbench/backend/spawn_boundary.py#L159)
max_attempts introduces more samples. Even if a global seed is set, objects are not guaranteed to be in the same position (https://github.com/stepjam/RLBench/blob/master/rlbench/backend/scene.py#L127)\

By using np.random.Generator, we can have a seed for a Scene instance, allowing potentially allowing more determinism. https://github.com/Farama-Foundation/Gymnasium/blob/main/gymnasium/utils/seeding.py

xsw1208 commented 3 months ago

Great, I also find that even if I have set global random seed, the boundary and object position varies.