AhernResearch / grava

GRAVA 
 Grounded Reasoning-to-Action Representation and Learning for Autonomous Driving 
 Xiao Liu 1,3 ·
 Haoyu Li 1,2 ·
 Jianghao Leng 1,3 ·
 Lin Wang 2 ·
 Chao Sun 1,3,* 
 1 Beijing Institute of Technology    2 Nanyang Technological University
 3 Shenzhen Jiguangzhijie Technology Co., Ltd.    * Corresponding author
 News ·
 Overview ·
 Highlights ·
 Method ·
 NAVSIM Examples ·
 Internal Examples ·
 Ecosystem ·
 Roadmap ·
 Citation 
## 📰 News
- 🚀 **[2026-10-04]** We’ve released grava-train to equip VLAs with spatial understanding and high-quality causal reasoning, alongside efficient training infrastructure for autonomous driving models.
- 🚗 **[2026-10-04]** grava-sim-engine is now available: our simulation and reward engine for training driving models with reinforcement learning.
## Overview
GRAVA is a vision-language-action framework that keeps physical scene evidence
connected to reasoning and executable driving behavior. Its **Grounded
Reasoning-to-Action (GRA)** representation binds linguistic references to visual
regions and ego-centric physical states, organizes object interactions and
driving decisions in a trajectory-anchored graph, and serializes the resulting
reasoning together with a compact planner action in one autoregressive stream.
The project covers the full research workflow: grounded data construction,
GR-NavSim, progressive model training, executable trajectory generation, and
NAVSIM evaluation.
## Highlights
- **Grounded reasoning that reaches the action.** Visual references and physical
 quantities remain attached to the interactions and decisions used for planning.
- **One autoregressive reasoning-to-action interface.** A single VLM generates
 grounded reasoning followed by an Executable Planner action, which is decoded
 deterministically into a continuous trajectory.
- **GRA-aligned data and learning.** Forward scene grounding and backward
 trajectory anchoring produce consistent cognition and planning supervision;
 progressive training combines grounded pre-training, planner warm-up, verified
 self-distillation, and Active RL.
## Method
GRAVA constructs a trajectory-anchored **Grounded Reasoning-to-Action (GRA)**
graph through forward scene grounding and backward trajectory anchoring, linking
visual references and physical states to object interactions, driving decisions,
and planner actions. The action-relevant subgraph is serialized as grounded
reasoning and followed in the same autoregressive sequence by an **Executable
Planner** action over STOP, CRAWL, CURVE, or CRUISE, which a fixed geometric
decoder converts into a continuous trajectory. Training combines graph-derived
GRA pre-training, planner warm-up, verified self-distillation, and Active RL on
recoverable scenes using trajectory-level driving rewards.
 GRAVA framework. An offline agentic pipeline constructs GRA supervision; a single VLM then learns to generate grounded reasoning followed by an executable planner action. 
## NAVSIM qualitative examples
 Grounde