Practical overview
When it fits
Use Pcla when the task involves evaluating pretrained CARLA-related driving agents, comparing their behavior in simulation, or connecting a supported agent to sensor data through the PCLA bridge.
Who it is for
- Autonomous-driving researchers comparing pretrained agents.
- Developers testing agents in CARLA scenarios.
- Teams investigating adversarial attacks against autonomous-driving agents.
- Builders exploring deployment of supported agents through the PCLA bridge.
Documented uses
- Evaluate supported autonomous-driving agents in CARLA.
- Compare multiple agents or training seeds within a common framework.
- Test agents with adversarial attacks.
- Retrieve an agent’s computed CARLA movement action for a custom application.
- Route external sensor data to a supported agent without CARLA through the bridge.
Requirements & limits
The repository is public and uses the Apache-2.0 license. The CARLA workflow requires local simulator, CUDA, PyTorch, Conda, assets, and pretrained weights.
- CARLA is required for the main simulator workflow, while the PCLA bridge is intended for simulator-free sensor-data integration.
- The README lists 41 supported agents, 27 additional training seeds, and an Autoware bridge.
- The project reports testing on Linux Ubuntu 22 and CARLA 0.9.16, with a note that CARLA 0.9.15 is more compatible with the agents.
- Running multiple vehicles with different agents requires high graphical memory.
- LMDrive has additional dependency and setup requirements.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up Pcla for CARLA-based agent testing
Prepare the documented Pcla environment and assets for testing supported autonomous-driving agents in CARLA.
Before you start
- A Linux Ubuntu 22 system, as reported for project testing.
- The CARLA simulator.
- CUDA and PyTorch.
- Conda.
- A CUDA-matched installation of torch-scatter and the CUDA Toolkit.
Steps
Clone the Pcla repository and enter its directory.
Create and activate the Conda environment from environment.yml.
Run the asset-download script.
Check the CUDA version with the included script, then install torch-scatter and the CUDA Toolkit for that version.
Download the required pretrained weights using the repository’s weight-download script.
Expected outcome
The repository and its documented environment, assets, and pretrained weights are prepared for the subsequent CARLA usage workflow.
Guide limitations
- The README reports testing with CARLA 0.9.16 and notes that CARLA 0.9.15 is more compatible with the agents.
- LMDrive requires additional Vision Encoder, LAVIS, and ftfy setup.
- The setup instructions do not establish that every listed agent works identically across environments.
Prepare a simulator-free PCLA bridge workflow
Use the documented bridge path when sensor data should reach a supported Pcla agent without requiring the CARLA simulator.
Before you start
- A supported Pcla sensor agent.
- A source of sensor data.
- The PCLA bridge repository components.
Steps
Open the PCLA bridge materials linked from the main repository.
Configure the bridge for the selected Pcla sensor agent.
Connect the available sensor-data source to the bridge.
Route the agent output to the intended use case.
Expected outcome
The documented bridge architecture provides a path for driving a Pcla sensor agent from external sensor data without CARLA.
Guide limitations
- The supplied README describes the bridge at a high level and points to a separate bridge directory for detailed usage.
- Vehicle deployment is described by the project, but this setup outline does not establish hardware compatibility or operational safety.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 4 October 2026
- Original website title
- MasoudJTehrani/PCLA · Official homepage title · retained in its source language