Skip to content

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.

  • Bicycle5D

    bicycle5d

    A 5-D bicycle drives to a goal through circular obstacles. avoid sits still; reach-avoid reaches from anywhere. Includes the learned value-function certificate and a PPO-vs-SAC value comparison.

    SafetyPPO1P · SafetySAC1P · ReachAvoidPPO1P · ReachAvoidSAC1P

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.