New Default Robot Simulator Environment

[mmmarinho]

@dqrobotics/developers and community,

Since the times of V-REP, CoppeliaSim has been the de facto simulator environment in dqrobotics, where the interface for at least 6 years was embedded in the main package for Python.

Recently with a monumental effort by @juanjqo and myself we have migrated to the zmq version of the API. Unfortunately we are still having issues such as

• API functions being blocking ([BUG] Performance issues with ZMQ cpp-interface-coppeliasim-zmq#4),
• Loss of compatibility between versions ([BUG] All methods of the experimental class related to the engine parameters settings are not working in CoppeliaSim 4.8.0-rev0 cpp-interface-coppeliasim-zmq#5)
• The EDU version is requesting a license, and because license request is blocked in previous versions, this might force us to update.
• Even the previous Player version has been renamed and is no longer available easily, meaning that we can't even be guaranteed to play scenes.

In conclusion, it might be time to choose another system as the default simulation environment. We don't need to drop support for the newly developed interfaces we made, but I don't see this situation being sustainable.

I'm adding choices in the order of the link below (CoppeliaSim is close to the end, otherwise it's in order).
https://github.com/knmcguire/best-of-robot-simulators

Please let me know your thoughts.

[juanjqo]

Hi @mmmarinho,

Overall, my experience with CoppeliaSim in the last months has been similar to the points you mentioned. I definitely would like to check some alternatives, even if we continue the CoppeliaSim support. Specially Webots, which I never used, and unlike O3DE, it is also available for macOS.

Kind regards,

Juancho

[ffasilva]

Hi @mmmarinho,

The motivations for seeking a new simulator seem reasonable.

I was trying to come up with some metric of comparison and have been focusing on the physics engines. From what I've read so far, MuJoCo seems to be the fastest and most accurate for multibody simulations [1]. That said, unfortunately, the comparison is from 2015, and much might have changed in the engines' performances. Also, the most recent general comparison (although AI-focused) I've found [2] classifies MuJoCo's documentation as "hard to navigate."

In contrast, Webots, mentioned by Juancho, seems to have better documentation. However, it uses the ODE as its physics engine, which in turn, has a "poorly structured documentation" [2].

Thus far, I still have no great way of choosing between our options, but I think we should start by ranking some basic requirements/metrics for that. The performance of the physics engine and the quality of the documentation are the ones I was checking. Any good ideas of what else we could compare?

Kind regards,
Frederico

[1] Erez, Tom, Yuval Tassa, and Emanuel Todorov. ‘Simulation Tools for Model-Based Robotics: Comparison of Bullet, Havok, MuJoCo, ODE and PhysX’. In 2015 IEEE International Conference on Robotics and Automation (ICRA), 4397–4404. Seattle, WA, USA: IEEE, 2015. https://doi.org/10.1109/ICRA.2015.7139807.
[2] Kaup, Michael, Cornelius Wolff, Hyerim Hwang, Julius Mayer, and Elia Bruni. "A review of nine physics engines for reinforcement learning research." arXiv preprint arXiv:2407.08590 (2024).

[mmmarinho]

Hi @juanjqo and @ffasilva,

Thanks for the replies.

We can start thinking of a list of requirements. We can start from the basics and then list what is needed. What I think we need to fill is the gap between a plot and something like IsaacSim. In my view, it can be something with a reasonable engine behind it, but for more accurate stuff, we could rely on other systems.

Examples to get the discussion going

• Multi-platform: Linux, MacOS, Windows
• Free and easy to install
• Can run on a regular laptop without a dedicated GPU (e.g., basic need is OpenGL).
• Covers all methods previously allowed by DQ_CoppeliaSimInterface
• C++ API
• Does not change the API every year
  among others

What other points would you like to add?

[bvadorno]

Hi @dqrobotics/developers,

The following points are also important:

• Large user base (it ensures longevity in the software usage).
• Most of our robots are available. Alternatively, the procedure for adding new robots is easy and quick.
• It has features that contemplate anticipated future use, such as contact-based manipulation.

Kind regards,
Bruno