Repository navigation
Consistent resetting through np.random.Generator #232
Description
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
Activity
xsw1208 commentedon Jun 23, 2024
Great, I also find that even if I have set global random seed, the boundary and object position varies.
vonHartz commentedon Jul 30, 2025
Is this still active?
akushonkamen commentedon Oct 17, 2026
Taking this one. Plan (minimal, backward-compatible): add an optional
np.random.Generator(orRandomState)rngthat is threadedEnvironment(rng=...) -> Scene(rng=...) -> SpawnBoundary(boundaries, rng=...), so reset-time boundary sampling no longer depends on global numpy state; whenrng=Nonethe behavior stays exactly as today (np.randomglobal). The global-np.randomdraws are atrlbench/backend/spawn_boundary.py:72-83(_get_position_within_boundary),spawn_boundary.py:103(rotation) andspawn_boundary.py:159(choice), consumed viaScene.init_episode/_place_task(rlbench/backend/scene.py:127-146, 529-537) which is what makes repeated resets non-reproducible even with a fixed global seed whenmax_attemptsretries occur.Will include a unit test for
SpawnBoundarydeterminism (same seed => identical placements, different seed => different) plus unchanged fallback behavior.Disclosure: I'm an AI-assisted contributor (account used by an automated coding agent); the fix and tests are prepared with AI assistance and reviewed by me before submission.
akushonkamen commentedon Oct 17, 2026
Opened PR #300 with a fix: an optional
rng(e.g.np.random.default_rng(seed)) is threadedEnvironment → Scene → SpawnBoundary, so reset-time placement sampling is reproducible independent of the global numpy state; the default (rng=None) keeps the existing global-np.randombehavior for all current callers. Includes CoppeliaSim-free regression tests.