Implementation:Haosulab ManiSkill BatchedRNG

Knowledge Sources	Haosulab_ManiSkill
Domains	Robotics, Simulation, Randomization
Last Updated	2026-02-15 08:00 GMT

Overview

Concrete tool for managing a batch of independent random number generators to ensure reproducible randomization across CPU and GPU simulators.

Description

The BatchedRNG class extends np.random.RandomState to manage a collection of independent random number generators (one per parallel environment). This ensures that randomization behavior is identical between CPU and GPU simulation backends when given the same seeds.

Key design:

• Each parallel environment gets its own np.random.RandomState instance, stored in the rngs list.
• from_seeds() class method creates the batch from a list of integer seeds.
• from_rngs() class method creates from pre-existing RNG instances.
• Indexing with [] returns a sub-batch or individual RNG.
• Assignment with []= replaces RNGs at specific indices.
• Any method call (e.g., .uniform(), .normal()) is automatically delegated to each individual RNG and the results are stacked into a numpy array. This is achieved via __getattribute__ override.

Usage

Used internally by ManiSkill environments for seeded randomization during reset. Particularly important for ensuring that CPU-based and GPU-based environments produce identical random configurations from the same seed.

Code Reference

Source Location

• Repository: Haosulab_ManiSkill
• File: mani_skill/envs/utils/randomization/batched_rng.py

Signature

class BatchedRNG(np.random.RandomState):
    def __init__(self, rngs: list): ...

    @classmethod
    def from_seeds(cls, seeds: list[int], backend: str = "numpy:random_state") -> "BatchedRNG": ...

    @classmethod
    def from_rngs(cls, rngs: list) -> "BatchedRNG": ...

    def __getitem__(self, idx: Union[int, list[int], np.ndarray]) -> Union["BatchedRNG", np.random.RandomState]: ...
    def __setitem__(self, idx, value) -> None: ...

Import

from mani_skill.envs.utils.randomization.batched_rng import BatchedRNG

I/O Contract

Inputs

Name	Type	Required	Description
seeds	list[int]	Yes	List of integer seeds, one per parallel environment
backend	str	No	RNG backend type (default: "numpy:random_state")

Outputs

Name	Type	Description
BatchedRNG	BatchedRNG	A batched RNG that delegates calls to per-env RNGs
method results	np.ndarray	Stacked results from calling any RandomState method

Usage Examples

Basic Usage

from mani_skill.envs.utils.randomization.batched_rng import BatchedRNG

# Create batched RNG for 4 parallel environments
rng = BatchedRNG.from_seeds([42, 43, 44, 45])

# Generate uniform random numbers for each env
values = rng.uniform(low=0, high=1, size=(3,))  # shape (4, 3)

# Index into a subset
sub_rng = rng[[0, 2]]  # BatchedRNG with 2 RNGs

Related Pages

• Environment:Haosulab_ManiSkill_Python_SAPIEN_Core
• Haosulab_ManiSkill_CommonUtils