Cannot run two scenarios back to back

[abol-karimi]

Using Scenic's Python API, I cannot run a scenario immediately after another. For the second simulation, I get the following error:

  File "tests/scenic2/loop.py", line 28, in <module>
    res2 = simulate()
  File "tests/scenic2/loop.py", line 19, in simulate
    sim_result = simulator.simulate(
  File "/home/carla/Scenic/src/scenic/core/simulators.py", line 76, in simulate
    simulation.run(maxSteps)
  File "/home/carla/Scenic/src/scenic/core/simulators.py", line 216, in run
    raise InvalidScenarioError(f'agent {agent} tried incompatible '
scenic.core.errors.InvalidScenarioError: agent unnamed Car (139820564322960) tried incompatible  action(s) (<scenic.domains.driving.actions.RegulatedControlAction object at 0x7f2a8315ad60>,)

Platform:

• Carla 0.9.14 docker
• Python 3.8.17 compiled from source
• Scenic 2.1.0b4

Here is the loopy.py script:

#!/usr/bin/env python3.8
import scenic

seconds = 20
timestep = .05
render = False

def simulate(scenic_file):
    scenario = scenic.scenarioFromFile(
                    scenic_file,
                    params={'timestep': timestep,
                            'render': render,
                            },
                    cacheImports=False
                    )
    scene, _ = scenario.generate(maxIterations=1)
    simulator = scenario.getSimulator()
    sim_result = simulator.simulate(
                        scene,
                        maxSteps=int(seconds // timestep),
                        maxIterations=1,
                        raiseGuardViolations=True
                        ) 
    return sim_result

res1 = simulate('tests/scenic2/dynamic1.scenic')
res2 = simulate('tests/scenic2/dynamic2.scenic')

and the two scenic files dynamic1.scenic and dynamic2.scenic specify the same scenario:

param carla_map = 'Town05'
carla_map = globalParameters.carla_map
param map = f'/home/carla/CarlaUE4/Content/Carla/Maps/OpenDrive/{carla_map}.xodr'
model scenic.simulators.carla.model

p0 = network.intersections[0].incomingLanes[0].centerline.pointAlongBy(0)

ego = Car at p0, facing roadDirection,
  with behavior FollowLaneBehavior(4)

[Eric-Vin]

Hi Abolfazl,

That looks like one of the bugs in this issue, which was fixed in this PR. The good news is that the fix should have been backported to 2.x and made it into 2.1. Can you try updating to Version 2.1 of Scenic (which should be compatible with what you have right now) and see if that resolves the issue?

[abol-karimi]

Thanks for the quick reply! The minimal example above works fine with v2.1. However, my actual scenic code (which is a modular scenario) still triggers the same error.

[Eric-Vin]

Can you provide your code that reproduces it? Then we can investigate why things are still crashing.

[abol-karimi]

I see that scenic.simulators.carla.actions keeps a reference to scenic.simulators.carla.model in order to access the Vehicle and Pedestrian classes, which seems similar to the bug that was fixed.

[dfremont]

You're right, good catch! We'll think about how to fix this.

[dfremont]

@abol-karimi Hopefully this is now fixed in main; please give it a try and let us know.

[abol-karimi]

I applied your changes to v2.1.0 and it works, thank you!

I had tried porting my project to Scenic 3 but I noticed a significant slow down.

[Eric-Vin]

Glad to hear it's working now. We'll back-port the fix to 2.x!

How much of a slowdown are you experiencing? We're definitely trying to avoid a significant slowdown for 2.0 scenarios in Scenic 3.0. Could you share your scenario with us so that we can profile it and try to ameliorate the slowdown?

[abol-karimi]

That would be great! Since we want to publish the results, the code is in a private repository. I'll add you to the repo and will show you how to run it.

[Eric-Vin]

Sounds good! If you'd like you can email me at evin {at} ucsc.edu

[Eric-Vin]

The patch (and several others) have been back-ported to 2.x in this PR: https://github.com/BerkeleyLearnVerify/Scenic/pull/173 .

Please let me know if there's any other issues, but I'll go ahead and close the issue for now.