Privately Owned Vehicle Work Group Meeting - 2026/05/18 - Slot 1&2

[TranHuuNhatHuy]

Agenda

• Milestones
• AutoDrive model design and development
• SodaSim integration update
• Camera calibration full implementation
• Development of robust photorealistic scene converter

Milestones

VisionPilot 1.0 was originally planned for release by the end of June. The team is now aiming to accelerate the timeline and release it sooner (hopefully mid-May).

New VisionPilot Architecture

AutoDrive Model Design and Development

@pranavreddy23

• Currently developing model base interfaces for VisionPilot 1.0 (Reference).

@atanasko

• Currently developing camera calibration module for VisionPilot 1.0. PR open soon.

Camera input/output bridges

@TranHuuNhatHuy

• ROS2 => OpenCV bridge: merged
• V4L2 => OpenCV bridge: merged
• [WIP] Image => WebRTC visualization streaming module.
• [WIP] Main visualization pipeline

Blind spot detection

• Currently being developed by @bilyas2 first PR

CAN interface

• @tranchithanh2000 and Ram (GitHub ID?) are working on it.

Robust Photorealistic Scene Converter

• Objective
  • Develop a lightweight, fast video style transfer pipeline to convert CARLA simulator images into images resembling the Waymo Open Dataset.
• Key Requirements
  • Runs on modest GPUs
  • Achieves ≥10 FPS at 256 × 512 resolution
  • Preserves semantic consistency (e.g., cars remain cars)
  • Produces high-quality photorealistic results
• Current Progress
  • Anthony and Thomas
  • Implementing CARLA2Real as an off-the-shelf solution:
  • https://github.com/stefanos50/CARLA2Real
• Next steps:
  • Evaluate performance
  • Optimize and fine-tune for project needs

Meeting Recording

• Zoom meeting recording available upon request
• Contact @m-zain-khawaja for access

[TranHuuNhatHuy]

Slot 1 meeting log:

• Current progress:

  • Anthony and Thomas are currently inspecting the NVIDIA's Physical AI Autonomous Vehicle and use it to train the Pix2Pix HD.
    • There are some concerns regarding traffic sign quality retaining, but let's try some approaches.
  • Then, they hope to integrate this model to AlpaSim.

• Next steps:

  • Finish the training of Pix2Pix HD via the NVIDIA dataset.
  • Attempt to reach a fair-enough quality of photorealistivity.
  • If above goal is achieved, try to retain the traffic sign quality using the semantic masking approach discussed in the meeting.

[TranHuuNhatHuy]

@TranHuuNhatHuy chaired Slot 1 meeting.

[ThomasLin617]

thomaslin617 attended Slot 1 meeting.

[Anthony1011]

Anthony attended Slot 1 meeting.

[atanasko]

attend slot 2

[pranavreddy23]

@pranavreddy23 slot 2