robosuite

Choose a release and a rendering route

The installation guide officially supports macOS and Linux and recommends an isolated Python environment, with Python 3.10 in its Conda example. The selected package metadata requires MuJoCo 3.3.0 or newer, pins mink to 0.0.5 and lists additional numerical and rendering dependencies. Its broad Python metadata alone does not establish that every Python 3 version works with the full dependency set. Record the actual resolved versions for an experiment. [S005] [S006]

The documented first check is the random-action demo. On macOS, the default MuJoCo viewer uses mjpython rather than the ordinary Python launcher. The same guide includes Windows troubleshooting, but its official support statement names macOS and Linux. Optional devices and integrations need additional dependencies. This review did not install the package or run its demo. [S005]

Observations and actions are configuration choices

The environment guide constructs a TwoArmLift task with Sawyer and Panda robots and a composite controller loaded through load_composite_controller_config. Robots, grippers, robot arrangement and controller configuration can all change. A controller change can also change the action space; an action vector from one configuration should not be assumed to mean the same thing in another. [S007]

For a first integration, use the environment's action_spec bounds and the documented reset/step loop. Observations are dictionaries rather than one universal flat vector. They can contain proprioception, object state, RGB images and depth, depending on the enabled modalities. Record those choices with the policy configuration so a comparison uses the same information at training and evaluation. This recording recommendation is an editorial reproducibility step. [S007]

The image-based example enables off-screen rendering and camera observations while disabling the on-screen renderer. A headless process therefore still needs the rendering route appropriate to its selected observations. The default image convention is OpenGL-style, described upstream as flipped; IMAGE_CONVENTION can switch to OpenCV-style output. That difference matters when reusing camera preprocessing. [S007]

Success and termination are different signals

The guide distinguishes sparse task-success rewards from shaped rewards. It also states that reaching the success criterion does not terminate the episode: environments continue to the configured horizon. A comparison that treats the returned termination signal as a success flag would therefore measure a different quantity. Preserve the reward setting, success definition, control frequency and horizon with evaluation results. No success rate or timing measurement was produced here. [S007]

Custom composite controllers require more than a configuration name. The extension guide subclasses a composite controller, registers it, implements its behavior, imports the class and supplies matching configuration. Its third-party example illustrates the route; it is not evidence that an arbitrary external controller is compatible. [S008]

The root license is MIT and separately identifies included MuJoCo portions under Apache-2.0. It does not establish blanket rights to every external model, renderer or dataset a user might select. This entry covers the simulation framework and its random-action setup route, without downloading a policy or demonstration dataset or claiming hardware transfer. [S003] [S005]