Stable-Envs
Fast and Reproducible Environments with Controllable Variations
stable-envs is a library of batched simulation environments that keep the full simulator state, so you can restore any step exactly and render images from it later, on CPU or GPU.
Store the state, replay or branch from any step, and draw frames from it later. 80 environment classes in 11 families share the same API. Policies, datasets and training live in stable-worldmodel, which builds on stable-envs.
Install¶
uv add stable-envs
pip install stable-envs
The core install runs CartPole, Acrobot and most navigation classes. Install lists the extras for the other families.
Example¶
import torch
import stable_envs as snv
with snv.CartPole(num_envs=8, device="cpu") as env:
env.reset(seed=0)
actions = torch.randint(0, 2, (8, 10, 1)) # (B, T, A): ten steps for each environment
out = env.step(actions, frames=None) # step without rendering
print(out["states"].batch_size) # torch.Size([8, 10])
frames = env.render(out["states"][:, 4]) # render the state after five steps
print(frames.shape, frames.dtype) # torch.Size([8, 400, 600, 3]) torch.uint8
step returns the state after every step, and render draws a batch of frames from any of them.
Next steps¶
| If you want to | Read |
|---|---|
| run your first environment | Quickstart |
| install the extras for a family, or set up a GPU | Install |
| store states in a dataset and render them while you train | Store states, render later |
| randomize colors, lighting or physics across a batch | Randomize appearance and physics |
| write an environment of your own | Write your own environment |
| find an environment and see what it looks like | Environments |
| look up a function or class | API |