At the core of Isaac Gym, “Physx” is basically the physics engine of computer games, isn’t it?
Not “gpu pipeline”.
Computer games do not require rigor in mission-critical issues.
Therefore, it can be inferred that serious problems will occur if Isaac Gym is applied to mission-critical issues.
“MuJoCo” is an abbreviation for “Multi-Joint dynamics with Contact” and is a physics engine suitable for articulated dynamics simulation. It’s free at this time.
IsaacGym is compatible with or has something in common with “MuJoCo”.
“Open AI” and “DeepMind” use MuJoCo ".
Jensen Huang said he would focus on robotics, right?
I predict that IsaacGym will be forced to use “MuJoCo” in the future. Is there such a possibility?
If IsaacGym prepares its own physics engine, shouldn’t it be developed from the beginning for articulated dynamics simulation instead of diverting “Physx (COMPUTER GAME)”?
Because it would make Isaac Gym completely lose its speed performance advantage and become no different from any other simulation environment.
The advantage of Isaac Gym is that both physical simulation and neural network training reside on the GPU, without the computational cost of communicating with the CPU, which makes the cost of parallel training very cheap, so as to accelerate the training speed of reinforcement learning.
If instead Mujoco is used for physical simulation, it has to communicate with the CPU again, which makes Isaac Gym meaningless.
From another point of view, this is incompatible with Nvidia’s software ecology. Isaac Gym will be the API of Isaac Sim in the future, and Isaac Sim is a part of Omniverse. They are all based on PhysX.
Of course, I don’t think that the current “MuJoCo” can be applied to “Isaac Gym” as it is. It is theoretically possible to change the physics method of “Physx” to the physics method of “MuJoCo”. It doesn’t matter whether it is GPU pipline or CPU pipeline.
NVIDIA mentioned in the topic below that it doesn’t have much of an advantage for Isaac Gym. I’m a minimalist so I don’t need a lot of features. I think the current stand-alone form is good.
I agree Issac Gym is enough if you just want to validate reinforcement learning algorithms. But I think Isaac Sim is more thinking about the Sim2Real problem.
When you observe the state of the environment, reading it directly from the physical engine is completely different from reading it through a sensor. This difference is especially evident in visual or SLAM sensors, I think Isaac Sim can better simulate these complexities, which helps you to adjust your model in time.