AhernResearch/grava

What it solves

GRAVA addresses the challenge of creating autonomous driving models that can effectively link visual scene understanding to concrete driving actions. Traditional models often struggle to maintain a clear connection between what the car "sees" and the specific decisions it makes, leading to gaps in reasoning and planning.

How it works

GRAVA uses a "Grounded Reasoning-to-Action" (GRA) representation. It builds a trajectory-anchored graph that connects visual objects and physical states to driving decisions. A single Vision-Language Model (VLM) processes this information to generate a sequence of reasoning followed by an "Executable Planner" action. This action is then converted into a continuous driving trajectory using a fixed geometric decoder. The training process includes grounded pre-training, planner warm-up, self-distillation, and reinforcement learning.

Who it’s for

This project is for researchers and developers working on autonomous driving systems who need an end-to-end framework for training models that combine high-level reasoning with precise, executable trajectory planning.

Highlights

• Grounded reasoning that keeps visual references and physical quantities attached to planning decisions.
• A unified autoregressive interface that generates both reasoning and executable actions in one stream.
• A comprehensive research ecosystem including training infrastructure, a simulation and reward engine, and dataset support.
• Multi-stage training pipeline that incorporates Active RL and verified self-distillation.