Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/NVlabs/alpasim/llms.txt

Use this file to discover all available pages before exploring further.

The driver in AlpaSim is a policy for the ego vehicle that takes in sensor inputs and optional navigation commands, and outputs a trajectory for the ego vehicle to follow, along with other optional outputs, such as chain-of-causation reasoning text. The driver is specified by a pair of config files under src/wizard/configs/, one for the driver service itself, and one for the runtime (so that it provides the inputs required for the specific driver).

​
VaVAM (Default Driver)

The wizard uses VaVAM as the default driver. This is a vision-based action model that learns to drive from video inputs.

​
Using VaVAM

To explicitly define the VaVAM driver config: 
alpasim_wizard +deploy=local wizard.log_dir=$PWD/tutorial_vavam driver=[vavam,vavam_runtime_configs]

​
VaVAM Configuration

VaVAM runs at 2Hz inference frequency with the following parameters:
  • Inference frequency: 2Hz (500ms between decisions)
  • Force GT duration: 3 seconds
  • Simulation steps: 100
  • Camera resolution: 1920x1080
  • Frame interval: 500ms (synchronized with inference)
The model weights are automatically downloaded using data/download_vavam_assets.sh and stored in data/vavam-driver/.

​
Alpamayo-R1

Alpamayo-R1 is a reasoning-capable autonomous driving model with 10 billion parameters that can provide chain-of-causation explanations for its driving decisions.

​
Model Download

First, download the model weights from HuggingFace: 
huggingface-cli download nvidia/Alpamayo-R1-10B
The wizard will use the HF_HOME environment variable to find the system HuggingFace cache (~/.cache/huggingface by default). If the model weights do not exist locally, the driver service will automatically download them, but the download may timeout, requiring you to re-run.
Alternatively, you can specify the path to the model directory by setting the model.checkpoint_path configuration field.

​
Running Alpamayo-R1

1

Run with Alpamayo-R1

Execute the wizard with the Alpamayo-R1 driver configuration:
alpasim_wizard +deploy=local wizard.log_dir=$PWD/tutorial_alpamayo driver=[ar1,ar1_runtime_configs]
2

Visualize Reasoning (Optional)

To visualize the predicted chain-of-causation reasoning, change the generated video layout:
alpasim_wizard +deploy=local wizard.log_dir=$PWD/tutorial_alpamayo driver=[ar1,ar1_runtime_configs] eval.video.video_layouts=[REASONING_OVERLAY]
The Alpamayo-R1 model is large (10B parameters). Please ensure that your GPU has sufficient capacity to run it.

​
Alpamayo-R1 Configuration

Alpamayo-R1 uses multiple cameras:
  • Front wide camera (120° FOV): 1080x1900 resolution
  • Front tele camera (30° FOV): 1080x1900 resolution
  • Left cross camera (120° FOV): 1080x1900 resolution
  • Right cross camera (120° FOV): 1080x1900 resolution
All cameras run at 10Hz (100ms frame interval) with synchronized inference.

​
Video Layout Options

You can choose which video layouts to render via eval.video.video_layouts.
The default layout shows:
  • BEV (Bird’s Eye View) map visualization
  • Camera view
  • Metrics table
alpasim_wizard +deploy=local wizard.log_dir=$PWD/tutorial driver=[vavam,vavam_runtime_configs] eval.video.video_layouts=[DEFAULT]

​
Transfuser (Provisional)

As an example for how to integrate a different driver model, AlpaSim provides a provisional integration for the Transfuser policy, specifically the Latent TransFuser v6 (LTFv6) model developed for NAVSIM.

​
Model Download

First, download the Transfuser model weights and config from HuggingFace: 
huggingface-cli download longpollehn/tfv6_navsim model_0060.pth --local-dir=data/drivers/transfuser/
huggingface-cli download longpollehn/tfv6_navsim config.json --local-dir=data/drivers/transfuser/

​
Running Transfuser

Run the wizard with the Transfuser configuration: 
alpasim_wizard +deploy=local wizard.log_dir=$PWD/tutorial_transfuser driver=[transfuser,transfuser_runtime_configs]

​
Log Replay Driver

If you would like to force the ego vehicle to follow its recorded trajectory, instead of following the predictions of a policy, you can configure the simulation in log replay mode. 
alpasim_wizard +deploy=local wizard.log_dir=$PWD/tutorial_replay \
  runtime.endpoints.physics.skip=true \
  runtime.endpoints.trafficsim.skip=true \
  runtime.endpoints.controller.skip=true \
  runtime.default_scenario_parameters.physics_update_mode=NONE \
  runtime.default_scenario_parameters.force_gt_duration_us=20000000
In log replay mode, the ego vehicle follows the exact trajectory from the original recording, which is useful for evaluation and debugging purposes.