lemketron
commented
3 years ago
- Apollo / SVL installed from (source or binary): Apollo installed from source via:
git clone https://github.com/ApolloAuto/apolloFollowing the markdown file from Tutorial: End-to-end video tutorial of LGSVL simulator with Apollo driving
Please note this video and related markdown document is from July, 2020 and refers to LGSVL Simulator 2020.06 used with a pre-release version of Apollo 6.0.
SVL installed from binary with steps from the installation guide
- Apollo version (3.5, 5.0, 5.5, 6.0):
This is good as SVL Simulator 2021.2 (or later) is what you should be using.
However, it looks like you are somehow using Apollo 5.0 from the simulator branch of the lgsvl fork of Apollo 5.0, rather than Apollo (upstream) 6.0 or master.
- Step: The tutorial indicates Dreamview should be started with
./scripts/bootstrap_lgsvl.sh, but my branch only hasbootstrap.shin the scripts folder so I start Dreamview with this. Dreamview loads and I selected the "Mkz Standard Debug" option from the setup drop down (different from the tutorial's selection). All of the modules outlined in the tutorial turn on except 'Perception' and 'Traffic Lights', which turn off shortly after toggling on.
This is because bootstrap_lgsvl.sh is the lgsvl variant of the bootstrap.sh script in the Apollo upstream repository. In the lgsvl fork of Apollo 5.0 it's simply called bootstrap.sh. The fact that you don't have bootstrap_lgsvl.sh indicates that you are are not using upstream Apollo. You can also confirm this by typing git remote -v in your apollo directory to see where it was cloned from (but the simulator branch is a clue that it is most probably the lgsvl fork of Apollo 5.0).
Hypothesis:
- GPU is too small 1050 Mobile vs 1080 recommended (could be the issue, not sure how to confirm)
Yes, this GPU is on the low end for running ONLY SVL Simulator with a small and simple map. If you are running any complex environments OR trying to run Apollo on the same machine, 4GB in the 1050 is just not enough memory. You can see that in nvidia-smi which will show you that the memory is consumed by the simulator, your browser, and likely not leaving enough memory for Apollo's prediction module (even if you're not planning to use Apollo perception).
- Nvidia drivers in the docker container don't match the drivers installed on the host (however, they are both at 460.80 with Cuda 11.2)
- Nvidia-docker is not installed in the apollo docker. The tutorial never indicated a step to install this package inside of the docker - but maybe there is some sort of GPU issue inside the container. (Not sure how to diagnose further)
It's been a while since I did a fresh checkout and build of Apollo but I don't recall any concerns with docker or nvidia-docker inside the apollo docker nor having to install anything there. The main concern is to have the correct software installed on the host.
- A docker group issue. The tutorial is clear that dev_start.sh should not be run with sudo. Even when running without sudo, the password is still requested and I've followed docker's post install steps as described above. (Not sure how to diagnose further)
The sudo password will still be requested when running dev_start.sh as a (non-root) user. This is normal and to be expected. You should not need to run dev_start.sh with sudo.
Summary:
I think you are trying to run Apollo 5.0, but are also trying to use "modular testing" to avoid needing to run Apollo obstacle perception (and traffic light perception). "Modular testing" requires (upstream) Apollo 6.0 or (better still) Apollo master.
Please clone (and re-build) Apollo master from https://github.com/ApolloAuto/apollo and refer to the following for more information: https://www.svlsimulator.com/docs/system-under-test/apollo-master-instructions/ https://www.svlsimulator.com/docs/tutorials/modular-testing/
schromotion
System information
Hardware, OS Platform and Distribution: Operating System: Ubuntu 20.04.2 LTS CPU: Intel Core i7-7700HQ CPU @ 2.80GHz GPU: NVIDIA Corporation GP107M GeForce GTX 1050 Mobile Kernel: Linux 5.8.0-59-generic Architecture: x86-64 Cuda: 11.2 Nvidia Drivers: 460.80
Docker: 20.10.7
Nvidia-Docker: 2.6.0
Apollo / SVL installed from (source or binary): Apollo installed from source via:
git clone https://github.com/ApolloAuto/apolloFollowing the markdown file from Tutorial: End-to-end video tutorial of LGSVL simulator with Apollo drivingSVL installed from binary with steps from the installation guide
Steps to reproduce the issue:
ISSUE: I'm trying to follow the tutorial in Tutorial: End-to-end video tutorial of LGSVL simulator with Apollo driving but hitting an issue when running Apollo connected to the SVL simulator where the "Perception" and "Traffic Light" modules keep automatically switching off.
Following the tutorial's linked markdown file, it begins to get interesting here:
sudo!)./docker/scripts/dev_start.shResult: The script still asks for the sudo password despite not running with sudo. I have followed the steps docker provides to run without sudo, and can successfully run thedocker run hello-worldstep at the end of those instructions on my machineThe output of this command gives
...
Step: Following the tutorial's instructions
nvidia-smiin the docker (afterdocker/scripts/dev_into.sh) yieldsStep: the build step yields a passing build
...
Step: Opening the cyber bridge gives warning not seen in the tutorial
Step: Cyber monitor and running sim bridge output
with a sim configuration of
Step: However, the Cyber Visualizer fails with
Step: The tutorial indicates Dreamview should be started with
./scripts/bootstrap_lgsvl.sh, but my branch only hasbootstrap.shin the scripts folder so I start Dreamview with this. Dreamview loads and I selected the "Mkz Standard Debug" option from the setup drop down (different from the tutorial's selection). All of the modules outlined in the tutorial turn on except 'Perception' and 'Traffic Lights', which turn off shortly after toggling on.Hypothesis:
Super excited about this simulator and learning more! Let me know if I had any issues or mistakes along the way or if there is a good path to finish the tutorial out from here.
ks