PSM24/alpamayo-r1-test
2
NVIDIA Alpamayo-R1-10B Inference Demo
A Gradio-based demo for NVIDIA Alpamayo-R1-10B, a Vision-Language-Action (VLA) model for autonomous driving.
Features
- Multi-Camera Input: Support for 4 camera views (front wide, front tele, left, right)
- Chain-of-Causation Reasoning: Interpretable decision-making traces
- Trajectory Prediction: 6.4-second future path prediction at 10Hz (64 waypoints)
- Interactive Visualization: Bird's eye view trajectory plots
Model Overview
Alpamayo 1 (formerly Alpamayo-R1) is NVIDIA's open 10B-parameter reasoning VLA model that:
- Takes multi-camera video input (4 cameras × 4 frames @ 10Hz)
- Outputs driving trajectories and reasoning traces
- Uses Chain-of-Causation (CoC) reasoning for interpretable decisions
- Built on Cosmos-Reason VLM backbone (8.2B params) + Action Expert (2.3B params)
Hardware Requirements
- GPU: Uses ZeroGPU (NVIDIA H200 with 70GB VRAM)
- Model Size: ~22GB (preloaded during build)
Usage
- Click "Load Model" to initialize
- Upload camera images (at least front wide camera)
- Optionally enter a driving command
- Click "Run Inference"
- View reasoning traces and trajectory visualization
License
- Model Weights: Non-commercial use only
- Inference Code: Apache 2.0
