Safety Gym is a highly configurable sandbox environment for testing safe RL and constrained RL algorithms. The flexibility is key for enabling rapid prototyping and testing of methods for safe reinforcement learning, which is a very active area of research in RL. It also has a really nice visualization component with full 3D rendering and simulation start/stop/step control that is useful for visualizing agents and their behavior.
However, Safety Gym is another example of a useful env that requires MuJoCo, which is a massive hindrance to the research community. My research group has been very interested in using this environment, and we've even attempted to acquire a MuJoCo license, but the maintainers of MuJoCo completely ignore us and don't respond to any of our emails. I think it would be a great boon to safe RL research to port this environment to an open source physics engine that doesn't require expensive licenses granted by 2-3 individuals who don't even respond to emails.
If other are interesting in porting this env, I can assist as well, I just don't have the bandwidth to do the entire port myself at my current position.
I would like to request a new addition to the suite of environments in pybullet-gym: OpenAI's Safety Gym:
https://openai.com/blog/safety-gym/
https://github.com/openai/safety-gym
Safety Gym is a highly configurable sandbox environment for testing safe RL and constrained RL algorithms. The flexibility is key for enabling rapid prototyping and testing of methods for safe reinforcement learning, which is a very active area of research in RL. It also has a really nice visualization component with full 3D rendering and simulation start/stop/step control that is useful for visualizing agents and their behavior.
However, Safety Gym is another example of a useful env that requires MuJoCo, which is a massive hindrance to the research community. My research group has been very interested in using this environment, and we've even attempted to acquire a MuJoCo license, but the maintainers of MuJoCo completely ignore us and don't respond to any of our emails. I think it would be a great boon to safe RL research to port this environment to an open source physics engine that doesn't require expensive licenses granted by 2-3 individuals who don't even respond to emails.
If other are interesting in porting this env, I can assist as well, I just don't have the bandwidth to do the entire port myself at my current position.