Autonomy Software WG Meeting 2022/03/29

[BonoloAWF]

Administrative

• Previous Meeting Minutes

Attendees

Chaired by Fatih

• Ryohsuke (AWF)
• Bonolo (AWF)
• Beyza (LeoDrive)
• Deepak
• Berkay (LeoDrive)
• Esteve (TierIV)
• Haoru (AWF)
• Hatem
• Lalith (TierIV)
• Fatih (AWF)
• Melike (LeoDrive)
• Mert (LeoDrive)
• Naoyori (TierIV)
• Kaan
• Ahmet
• Mehmet (LeoDrive)
• Shu-Kang (ITRI)
• Ambroise (Arm)
• Sena (LeoDrive)
• Zeynep
• Brian (ITRI)
• Yu (UTokyo)
• Goktug (LeoDrive)
• Henry (ITRI)
• Armagan (LeoDrive)
• Maxime (TierIV)
• Piotr (Robotec.ai)
• Xinyu (AWF)

Common Resources

• Bus ODD software architecture proposal
• Bus ODD architecture design discussions
• Bus ODD candidate simulator comparisons
• Bus ODD milestone development plan
• Autoware contributors list

Announcements

• If you haven't already added your Github username to the contributors list, please do so this week. Also remember to accept the Github invitation to join the developer team!

Autoware.Auto

• Review Cargo MS10 Cleanup & Technical Debt Issues
• Updates on bug fixes and documentation bugs
• Release of Autoware.Auto v2.0.0
• Preparing perception and planning packages for ADASTEC on-vehicle testing

Autoware.Universe

• msg architecture finalisation discussion
• Issues in the Bus ODD Project Milestones associated with a timeframe milestone

Discussion topics

Presentation: Scenario Runner Guide

Ryohsuke presented the scenario runner guide (see comments for presentation slides). Summary of the presentation:

• tool to test Autoware features using scenario files which communicate with the simulator
• tool can be augmented to CI to avoid weekly manual integration testing
• drive commands are provided to the MPC
• scenario editor is a GUI interface for creating scenarios (eg. set initial and target goals). Can export as YAML file which is fed into scenario runner which runs Autoware to execute the scenario.
• scenarios are defined by the ODD WG and we can use them for the cargo ODD and bus ODD
• generated scenario files are hosted in Gitlab (to be moved to Github soon)
• TierIV traffic simulator is the only one currently supported (no sensing/perception testing)
• other E2E simulators will be integrated
• see presentation for links to Documentation and tutorials to get started
• terminal output shows whether or not the scenario has passed

Hatem (H) - how do you tell the scenario you want to run Autoware with it? I get an error "waiting for Autoware to initialise" Ryohsuke (R) - in the launch file of the scenario editor the default refers to autoware_launch_file.
Bonolo (B) - has it been implemented to link scenario tester to PRs? R - Authors of PRs will attach a target scenario to the PR and reviewers can run the scenario to evaluate the PR. We need to update the templates to include this.
Haoru (H) - how familiar should people be with scenario tests when developing future Autoware packages? R - I would like scenario testing to be the default way of testing in Autoware. In the PR template we can have a simple instruction added on how to run scenario testing as a reviewer.

Proposed by Ryohsuke:

If we have time, I would like to discuss about:

1. Maybe fix the stopping problem of pos_db. #96 Very long discussion was done about the installation script whether we should use pip or apt for installing rocker
2. I would like to discuss what everyone thinks about models for machine learning based perception nodes. We have CenterPoint uploaded as GDrive file but it sometimes fail with the download using a download script (gdown). Should modelzoo only support TVM models? or can it support other models. Who would maintain the repository?

The rocker was previously installed using pip command but this was modified to be installed via apt. If anyone has a preference or experience with this please comment on the issue. The software architects will meet to discuss a framework for dealing with disagreements related to software development (usually voting works well).

Ryohsuke (R) - there are download bandwidth limitations which trigger additional costs per download. Center point model uses Tensor RT primers and would need to be converted to TVM in order to be included in the model zoo. Ambroise (A) - model zoo models are TVM based, additional work is required to support non-TVM models. R - Ambroise, do you know the setup for model zoo? A - Yes. Fatih (F) - for S3 bucket there is autoware.ai, autoware.auto, autoware-model-zoo, and autoware-gitlab-runner-cache. So far billing was around $400/month, there is more cost on the data transfer. Maybe we can use S3 bucket? A - S3 bucket for model zoo is for development only. Downloading models is disabled by default so it is lightly utilised since packages aren't actually using the models. This would significantly increase data transfer and associated costs. R - we should keep the same approach to auto disable downloads. We need to discuss further including moving the TVM pipeline to Universe.

Proposed by Fatih:

• Discussion about localization: Remove extra level of indirection from cuMemFreeHost #112

Universe has 2 packages for converting to Cartesian coordinates (eg. gnss_poser) which derives the PoseStamped and PoseWithCovarianceStamped msgs. GeographicLib provides a library for conversions which is easier to use. See: https://geographiclib.sourceforge.io/C++/doc/

Action Items

1. @mitsudome-r to contact NE Raptor developers to discuss https://gitlab.com/autowarefoundation/autoware.auto/AutowareAuto/-/issues/1102
2. @mitsudome-r to ask ODD WG to create a presentation that explains scenario testing with Autoware tools
3. @BonoloAWF to create labels for bus ODD, LTB and assign timeframe milestones to issues
4. @armaganarsln to create a Discussion for investigating radar in Autoware

Review of issues

Ryohsuke - if CI passes, the Autoware.Auto MRs should get merged and the release can be completed. We want ADASTEC to test Autoware.Universe rather than Autoware.Auto. I can list the limitations identified so we can come up with tests for Universe packages.

Fatih - developers should review semantic pull request docs which check if the name of the PR matches the format require. All comments should be signed off according to DCO. See: https://autowarefoundation.github.io/autoware-documentation/main/contributing/pull-request-guidelines/#dco. Ryohsuke - these rules exist to improve traceability of changes made.

[xmfcx]

Discussion about localization: #112

[mitsudome-r]

If we have time, I would like to discuss about:

1. Maybe fix the stopping problem of pos_db. #96 Very long discussion was done about the installation script whether we should use pip or apt for installing rocker
2. I would like to discuss what everyone thinks about models for machine learning based perception nodes. We have CenterPoint uploaded as GDrive file but it sometimes fail with the download using a download script (gdown). Should modelzoo only support TVM models? or can it support other models. Who would maintain the repository?