Research map / Vision-language and multimodal / Multimodal LLMs

Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Drive-R1 adapts a small InternVL2-4B driving VLM via two-stage SFT (3M QA pretraining, then fast-and-slow short/long CoT data) followed by GRPO with trajectory, meta-action, repetition and format rewards for reasoning-aligned motion planning.

Problem

VLM-based motion planners for autonomous driving under-use visual input (relying on history text priors) and their chain-of-thought reasoning is misaligned with the planned trajectories.

Approach

Combine domain-specific SFT with an adaptive fast-and-slow (short vs. long CoT) supervision strategy, then apply GRPO with a composite reward coupling trajectory accuracy and meta-action correctness so reasoning aligns with planning outcomes.

Key results

• Drive-R1 achieves average L2 of 0.31 m on nuScenes validation, versus 0.32 for EMMA and 0.33 for OmniDrive.
• Drive-R1 reports average collision rate of 0.09 on nuScenes, lower than RDA-Driver (0.10) and OmniDrive (0.30).
• On DriveLM-nuScenes ablation, adding meta-action reward and repetition penalty reduces average collision from 0.14 to 0.10.

Surprising findings

• A trajectory-only VLM performs comparably or better when visual inputs are ablated, indicating reliance on historical textual cues.
• Adding long CoT supervision in SFT degraded planning performance relative to direct trajectory supervision.

Limitations

• Small-capacity model limits ability to exploit complex reasoning; CoT supervision can hurt numerical planning.
• RL training becomes unstable with 24 rollouts beyond a certain number of steps.
• RP-CoT dataset is small (about 4k samples) and relies on synthetic ChatGPT/GPT-4o annotation.

Experimental setup

Datasets: nuScenes DriveLM-nuScenes self-collected driving QA (3M samples) RP-CoT (4,072 short/long CoT samples)

Benchmarks: nuScenes DriveLM-nuScenes

Metrics: L2 distance (m) collision rate

Baselines: ST-P3 UniAD UniAD-E VAD-E DriveVLM RDA-Driver OmniDrive EMMA

Builds on: GRPO DeepSeek-R1 InternVL2 GPT-Driver ST-P3

Closest papers by meaning

• $AutoDrive\text{-}P^3$: Unified Chain of Perception–Prediction–Planning Thought via Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.95)
• DriveAgent-R1: Advancing VLM-based Autonomous Driving with Active Perception and Hybrid Thinking (ICLR 2026 · similarity 0.93)
• AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.93)
• OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation (NeurIPS 2026 · similarity 0.93)
• LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.92)
• From Representational Complementarity to Dual Systems: Synergizing VLM and Vision-Only Backbones for End-to-End Driving (NeurIPS 2026 · similarity 0.92)