Environments¶
A growing set of small, self-contained reference environments that ship with
safety_sb3 (under safety_sb3.testing) — each trainable to convergence in
minutes on CPU, and each exercising the library on a problem where the right
behavior is obvious. They are the fastest way to see the algorithms work and to
validate your own setup against a known-good baseline.
For the mjlab / GPU robot environments (humanoid, quadruped, etc.), see robot-safety-sandbox and its own environment showreel.
Each page follows the same shape: task definition, the observation / action /
margin contract, a few lines to launch, and the expected result (GIF + value
figure). To add one, drop a module in safety_sb3/testing/ and a page here.
