The viewer currently hard crashes when the user script loops on the passive viewer handle without paying attention to `is_running`. This is considered user error, however this change makes the user script hang on the loop rather than hard crashing.
Startblock:
diffbase is submitted
PiperOrigin-RevId: 572336261
Change-Id: Id53126a158242b947584da1f946304ae8e792d4a
This change also requires user scripts to explicitly synchronize changes to physics state to the viewer. The Simulate class was reconfigured so that certain UI events are handled during this sync operation, outside of the render loop on the main thread. These correspond to operations that require access to the full mjModel/mjData.
To support other, more interactive operations (e.g. camera movements), a new mjvSceneState struct is introduced which captures only the portion of the physics state required for scene re-rendering. The mjvSceneState is updated from mjModel/mjData during the viewer sync operation, and is significantly cheaper than a full mj_copyModel and mj_copyData.
Fixes https://github.com/deepmind/mujoco/issues/796
PiperOrigin-RevId: 525723636
Change-Id: Id08d0210a2c067d5afe85e2bf104f276aeddd75e
Before this change, these functions can return before an internal call to `mj_forward`, which results in a race condition if the user script then tries to modify the physics state.
Fixes#783Fixes#790
PiperOrigin-RevId: 523214378
Change-Id: I8a4c6cef0f5d3ec781fd2c5cae3c20dd4b8be67f
The `launch_passive` function launches the GUI viewer in a non-blocking manner, allowing the Python script or REPL to continue execution. The viewer is automatically kept up to date with any subsequent modifications to mjModel and mjData.
Note that when run inside a REPL (including IPython), `launch_passive` is functionally identical to `launch_repl`.
On Linux and Windows, this is achieved by spawning a new thread and launching the GUI there.
On macOS, this is not possible as all Cocoa API calls must be made on the "macOS main thread", which is always the first thread launched in a process and carries the `com.apple.main-thread` dispatch queue. We also cannot simply trampoline from a Python script on the main thread into the user's script on a side thread because CPython's signal handler can only be installed on the "Python main thread". Putting the user's script in a side thread means that it cannot e.g. gracefully handle SIGINT by catching a KeyboardInterrupt exception. To work around this, we ship a custom Python launcher on macOS called `mjpython`. This launcher is a native binary that spawns a pthread and initialize the Python interpreter on that thread, thus allowing "Python main thread" and "macOS main thread" to represent two distinct threads. From Python's point of view, the "macOS main thread" is a secondary thread that runs a loop that continuously empties a Queue of (mjModel, mjData) and launches a viewer.
PiperOrigin-RevId: 517167868
Change-Id: Icac9d2126bbb4760d47e0b9300e0a979cffa4338
Copybara-Service