Reproduce “GRAVA: Grounded Reasoning-to-Action Representation and Learning for Autonomous Driving” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiao Liu, Haoyu Li, Jianghao Leng, Lin Wang, Chao Sun
GRAVA links grounded visual reasoning directly to executable driving actions by representing action-relevant objects, their physical states, interactions, and decisions in a shared reasoning sequence. On NAVSIM, the 8B model reaches a PDMS of 90.48 after active reinforcement learning, while improving long-tail key-object compliance from 77.4 to 92.3 and closed-loop driving score from 74.8 to 90.1 over an action-only variant.
Key insights
- Backward anchoring from expert trajectories helps distinguish objects that merely appear in the scene from objects that actually constrain the demonstrated maneuver.
- The executable planner provides a compact interface between language reasoning and continuous motion, while keeping action generation conditioned on the preceding grounded explanation.
- Intervention experiments support a causal connection between reasoning quality and planning: replacing weak reasoning with higher-reward grounded reasoning raises normalized PDMS from 0.438 to 0.823 and increases the reported win rate from 3% to 55%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.