MAGNIFIED: Google's RL fine-tuning cuts autonomous driving errors by 38%

Multi-modal Large Language Models (MLLMs) show promise for autonomous driving planning due to their semantic understanding and commonsense reasoning. However, traditional supervised fine-tuning (SFT) with next-token prediction objectives often fails to align with critical driving goals like safe spacing and off-road avoidance. To address this, a team of researchers (including Waymo) introduces MAGNIFIED, a reinforcement learning fine-tuning approach that teaches MLLMs to optimize for planning-specific rewards rather than just imitating token sequences. By mapping predicted tokens to vehicle trajectories and learning from rewards that penalize collisions and off-road incidents, MAGNIFIED directly improves real-world driving metrics.

The method was validated on the Waymo Open Motion Dataset using rasterized bird's-eye views and tokenized trajectory inputs. An initial SFT phase established a baseline, and subsequent RLFT with MAGNIFIED delivered a 10.5% reduction in overlap rate (collisions) and a 38.9% reduction in off-road rate. These gains underscore the potential of RL fine-tuning to produce MLLM-based driving agents that are more compliant, comfortable, and efficient. The work positions MAGNIFIED as a key step toward deploying LLMs in safety-critical autonomous vehicle systems.