Reset models and worlds similarly to Ctrl + R in Gazebo Classic

[diegoferigo]

I see a lot of new GUI-related issues to track new features, and that's great! There's one feature, though, that in Gazebo Classic was used extensively by pretty much all the people I know that's still missing: resetting a world with Ctrl + R.

Gazebo Classic had the following APIs:

• gazebo::physics::Model::Reset
• gazebo::physics::World::Reset

I never had a look at how it was implemented previously, what I want to add to this issue as a reminder is that it would be nice having a sort of callback that world and model plugins could use to reset their state.

A simple alternative would be adding a new reset component and ask plugin developers to always check if this component is present and take action. Both would equally work, at first thought I don't see major limitations of any of them. Maybe the callback would propagate better and more clearly.

Partially related to #113.

[chapulina]

Thank you for ticketing this!

Since this PR we already have a message interface to reset simulation time, for example:

ign service -s /world/shapes/control --reqtype ignition.msgs.WorldControl --reptype ignition.msgs.Boolean --timeout 5000 --req 'reset: {all: true}'

The way systems can react to this is by looking at UpdateInfo (not a callback or components though).

From the PR:

---

Systems can detect rewinds by checking:

if (_info.simTime == std::chrono::steady_clock::duration::zero())
if (_info.dt < std::chrono::steady_clock::duration::zero())

To detect pause:

if (_info.paused)

To detect time jump while paused (dt is zero during regular pause):

if (_info.paused && _info.dt != std::chrono::steady_clock::duration::zero())

---

So I think the main missing pieces are:

• Button / hotkey on the GUI to call the world control service and reset. This could be added to the WorldControl ign-gui plugin.
• Making each system react properly to jumps back in time. This is the tough part, especially for physics.

[diegoferigo]

Great, thanks for providing these details! If I understood well, there's already the initial support for something similar than Ctrl+R, actually more advanced since it allows selecting t=t_k instead of t=0 by default. However, so far there's no System yet (including Physics) that use this feature.

Few comments:

• While it is handy having a transport-based interface, a C++ API is still missing. I don't want to go much in detail into design choices (discussions about network transparency probably would require a separated issue) but in my opinion one of the selling points of Ignition Gazebo is the possibility of using it as a library without the need to use transport. In past few years other simulators like mujoco and pybullet got an remarkable user-base thanks to this feature and I hope that also Ignition Gazebo could reach a similar feature-set (and way beyond :)).
• Beyond time seeking, I still think that APIs like World::Reset and Model::Reset are quite useful.
• Regarding the reset of the Physics, I have the feeling that rolling back time (by issuing negative time steps) and resetting the state of the physics engine + aligning the ECM are two very different approaches that could be implemented separately. The former could involve a (backward?) integration, the latter just a reset of the state taking care of all the hidden states like integrator buffers in the case of PIDs (maybe the reset components of this PR could help). Using time = 0 maybe is easier since the state could be gathered from the SDF of models and world, a reset to a given time would require the logger being enabled.

[chapulina]

so far there's no System yet (including Physics) that use this feature.

That's true for live simulations, but the LogPlayback system supports it when playing back log files.

a C++ API is still missing

That's a valid point. Since time management happens outside of the ECM, I think a component approach would be tough to implement though. Probably an event like we have for pause would be more appropriate:

https://github.com/ignitionrobotics/ign-gazebo/blob/d7cd60a7ec8d5931f54d155894b66974dbcf78a2/src/SimulationRunner.cc#L103-L104

rolling back time (by issuing negative time steps) and resetting the state of the physics engine + aligning the ECM are two very different approaches that could be implemented separately

Fully agreed. The message interface already keeps those separate, with reset / rewind separate from seek.

[diegoferigo]

xref #1107