gym_solo VS pybullet-gym

You can find this simulation here: https://github.com/WPI-MMR/gym_solo

And when something like "RL/Fortnite Racing" comes out, with shared inventory, it may work as the pilot "beta test" of RL finally working in UE5. At least the game's physics engine (pybullet.org) is independent of the graphics. https://x.com/CactousMan/status/1690437803711320064

From what I know in pybullet you define the system of the environment loading an urdf file. So if you check the InvertedPendulum environment of pybullet I think you could create your desired urdf without a problem. I have not used mujoco, but it seems strange that it could not do what you want to do... Maybe there is a parameter of the joints or some kind of tolerance for the initial position?

So far I found PyBullet, RobotPy, RobotDK, SOFA, and some others, but I wonder if there is something that is comparable or better than Unity 3D for this specific use case.

yeah, i guess compared to modeling tools like tinkercad, blender is not intuitive and hard to learn. im not aware of any FEM with blender but there's bullet, basis of https://pybullet.org/. blender has an omniverse connector now so it could probably be used in a pipeline with nvidia's isaac. https://developer.nvidia.com/isaac-sim https://developer.nvidia.com/blog/introducing-omniverse-usd-support-for-blender/

Programming is also really important when learning anything about machine learning and I would start out with the OpenAI Gym and its derivative for learning some of the simpler algorithms on toy environments that doesn't require a whole lot of computing power. Then you can move on to the more hardware constrained simulators that I mentioned in the previous comment (MuJoCo, PyBullet, NVIDIA Isaac Gym).

shout-out to the open-source clones of the mujoco gym environments here