$ show NVlabs/alpamayo
Alpamayo 1 is a Vision-Language-Action (VLA) model designed for autonomous driving applications. It implements the Alpamayo-R1 architecture that combines a Qwen3-VL backbone for vision-language reasoning with a diffusion-based action expert for trajectory prediction. The system takes multi-camera video and egomotion history as inputs, generates Chain-of-Causation (CoC) reasoning traces, and outputs future trajectories (6.4s horizon at 10Hz). Key components include a unicycle kinematic model for action space conversion, CUDA graph acceleration for efficient repeated inference, and robust geometric transformations. While the repository is noted to receive only limited maintenance as newer versions exist, the current implementation demonstrates strong engineering practices with several reusable components.
The LLM identified 5 files worth highlighting for reuse:
Contains mathematical utilities for solving constrained optimization problems using Tikhonov regularization with first, second, and third order smoothing terms. Implements banded matrix construction for efficient solving of smoothing-regularized least squares problems. Demonstrates sophisticated numerical methods that could be reused in any project requiring smoothing or regularization of sequential data.
Implements a unicycle kinematic model with acceleration and curvature as control inputs for converting between trajectories and control signals. Includes sophisticated smoothing constraints (Tikhonov regularization) for velocity, acceleration, and jerk. Provides methods to estimate initial states, convert trajectories to actions, and convert actions back to trajectories. This self-contained component could be reused in other autonomous driving or robotics projects needing trajectory-control conversion.
Implements exact-shape CUDA graph replay for transformer-based diffusion experts to accelerate repeated inference. Captures and replays static computation graphs for specific input shapes while falling back to eager execution for new shapes. Properly handles prompt cache management and maintains compatibility with HuggingFace transformer interfaces. This optimization technique could be applied to many transformer models needing improved inference performance.
Provides robust geometric transformation utilities including conversion between rotation matrices and yaw angles, stable Gram-Schmidt orthonormalization, and 2D/3D rotation matrix conversions. Features a robust arctan2 implementation that avoids NaN outputs. These commonly needed geometric operations are implemented with attention to numerical stability and could be reused in any computer vision or robotics project.
Implements a multi-GPU friendly logger that prefixes log messages with process rank. Properly handles initialization to avoid duplicate log output when used with frameworks like Hydra. The RankedLogger adapter ensures logs from all processes include rank information while allowing rank-zero-only logging when needed. This reusable component solves a common problem in distributed PyTorch training and inference.